tutorials/Part 08 - Building a distinguisher.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Functional Encryption - Classification and information leakage\n",
    " \n",
    "We would like to have a measure of the leakage with cryptology notions. In particular, we will look at the notion of [advantage](https://en.wikipedia.org/wiki/Advantage_(cryptography)) in the distinction task.\n",
    "\n",
    "> An adversary's advantage is a measure of how successfully it can attack a cryptographic algorithm, by distinguishing it from an idealized version of that type of algorithm -- Wikipedia\n",
    "\n",
    "\n",
    "Let $Q$ be the real quadratic network, and let $Q_{ideal}$ be the real one with no collateral leakage (ie accuracy on the collateral task would be 50%). The adversary $F$ is a probabilistic algorithm given $Q$ or $Q_{ideal}$ as input and which outputs 1 or 0 (depending on the font). $F$'s job is to distinguish $Q$ from $Q_{ideal}$ based on making queries given data samples. We say: $Adv(F)=|\\Pr[F(Q)=1]-\\Pr[F(Q_{ideal})=1]|$\n",
    "\n",
    "Because the adversary is sharp, it will fix a digit and try to distinguish on this one the font used. We chose $6$ because we are fair and wanted a digit of average difficulty.\n",
    "\n",
    "So to be clear: when building the resistance, we must provide equal resistance against all kind of digits for the couple of fonts selected, but the adversary can be specialized in a single digit. Therefore the problem is a bit unbalanced and we expect worse results compared to the Part 6. Therefore, we already start with longer sabotage and attack phases to get a more realistic view of what we should expect."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 8. Collateral Learning assessed with Advantage\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "We will use the code directly from the repo, to make the notebook more readable. Functions are similar to those presented earlier."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Allow to load packages from parent\n",
    "import sys, os\n",
    "sys.path.insert(1, os.path.realpath(os.path.pardir))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "import random\n",
    "\n",
    "import numpy as np\n",
    "\n",
    "import torch\n",
    "import torch.nn as nn\n",
    "import torch.optim as optim\n",
    "import torch.nn.functional as F\n",
    "import torch.utils.data as utils\n",
    "\n",
    "import learn\n",
    "from learn import show_results\n",
    "# We now import from the collateral module\n",
    "from learn import collateral"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 58,
   "metadata": {},
   "outputs": [],
   "source": [
    "class Parser:\n",
    "    \"\"\"Parameters for the training\"\"\"\n",
    "    def __init__(self):\n",
    "        self.epochs = 10\n",
    "        self.sabotage_epochs = 10 # PUT 50 to get a proper resistance, but expect a long computation!\n",
    "        self.new_adversary_epochs = 50 \n",
    "        self.lr = 0.002\n",
    "        self.momentum = 0.5\n",
    "        self.test_batch_size = 1000\n",
    "        self.batch_size = 64\n",
    "        self.log_interval = 300"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "PRIVATE_OUTPUT_SIZE = 10\n",
    "N_CHARS = 10\n",
    "N_FONTS = 2"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 8.1 Loading $Q$ with resistance"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 110,
   "metadata": {},
   "outputs": [],
   "source": [
    "class CollateralNet(nn.Module):\n",
    "    def __init__(self):\n",
    "        super(CollateralNet, self).__init__()\n",
    "        self.proj1 = nn.Linear(784, 40)\n",
    "        self.diag1 = nn.Linear(40, PRIVATE_OUTPUT_SIZE, bias=False)\n",
    "\n",
    "        # --- FFN for characters\n",
    "        self.lin1 = nn.Linear(PRIVATE_OUTPUT_SIZE, 32)\n",
    "        self.lin2 = nn.Linear(32, N_CHARS)\n",
    "\n",
    "        # --- Junction\n",
    "        self.jct = nn.Linear(PRIVATE_OUTPUT_SIZE, 784)\n",
    "\n",
    "        # --- CNN for families\n",
    "        self.conv1 = nn.Conv2d(1, 20, 5, 1)\n",
    "        self.conv2 = nn.Conv2d(20, 50, 5, 1)\n",
    "        self.fc1 = nn.Linear(4 * 4 * 50, 500)\n",
    "        self.fc2 = nn.Linear(500, N_FONTS)\n",
    "        \n",
    "        \n",
    "        # FFN for families\n",
    "        self.lin3 = nn.Linear(PRIVATE_OUTPUT_SIZE, 64)\n",
    "        self.lin4 = nn.Linear(64, 32)\n",
    "        self.lin5 = nn.Linear(32, 16)\n",
    "        self.lin6 = nn.Linear(16, 8)\n",
    "        self.lin7 = nn.Linear(8, N_CHARS)\n",
    "\n",
    "    def quad(self, x):\n",
    "        # --- Quadratic\n",
    "        x = x.view(-1, 784)\n",
    "        x = self.proj1(x)\n",
    "        x = x * x\n",
    "        x = self.diag1(x)\n",
    "        return x\n",
    "\n",
    "    def char_net(self, x):\n",
    "        # --- FFN\n",
    "        x = F.relu(x)\n",
    "        x = F.relu(self.lin1(x))\n",
    "        x = self.lin2(x)\n",
    "        return x\n",
    "\n",
    "    def font_net(self, x):\n",
    "        # --- Junction\n",
    "        x = self.jct(x)\n",
    "        x = x.view(-1, 1, 28, 28)\n",
    "\n",
    "        # --- CNN\n",
    "        x = F.relu(self.conv1(x))\n",
    "        x = F.max_pool2d(x, 2, 2)\n",
    "        x = F.relu(self.conv2(x))\n",
    "        x = F.max_pool2d(x, 2, 2)\n",
    "        x = x.view(-1, 4 * 4 * 50)\n",
    "        x = F.relu(self.fc1(x))\n",
    "        x = self.fc2(x)\n",
    "        return x\n",
    "\n",
    "    def forward_char(self, x):\n",
    "        x = self.quad(x)\n",
    "        x = self.char_net(x)\n",
    "        return F.log_softmax(x, dim=1)\n",
    "\n",
    "    def forward_font(self, x):\n",
    "        x = self.quad(x)\n",
    "        x = self.font_net(x)\n",
    "        return F.log_softmax(x, dim=1)\n",
    "    \n",
    "    # We add the ability to freeze some layers to ensure that the collateral task does\n",
    "    # not modify the quadratic net\n",
    "    \n",
    "    def get_params(self, net):\n",
    "        \"\"\"Select the params for a given part of the net\"\"\"\n",
    "        if net == 'quad':\n",
    "            layers = [self.proj1, self.diag1]\n",
    "        elif net == 'char':\n",
    "            layers = [self.lin1, self.lin2]\n",
    "        elif net == 'font':\n",
    "            layers = [self.jct, self.fc1, self.fc2, self.conv1, self.conv2]\n",
    "        else:\n",
    "            raise AttributeError(f'{net} type not recognized')\n",
    "        params = [p for layer in layers for p in layer.parameters()]\n",
    "        return params\n",
    "\n",
    "    def freeze(self, net):\n",
    "        \"\"\"Freeze a part of the net\"\"\"\n",
    "        net_params = self.get_params(net)\n",
    "        for param in net_params:\n",
    "            param.requires_grad = False\n",
    "\n",
    "    def unfreeze(self):\n",
    "        \"\"\"Unfreeze the net\"\"\"\n",
    "        for param in self.parameters():\n",
    "            param.requires_grad = True"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 111,
   "metadata": {},
   "outputs": [],
   "source": [
    "def build_resistance(model, alpha=0):\n",
    "    \"\"\"\n",
    "    Perform a dual learning phase with sabotage\n",
    "    \"\"\"\n",
    "    args = Parser()\n",
    "    # We are just building Q so we don't need the 3rd phase, it will be assessed later!\n",
    "    args.new_adversary_epochs = 0\n",
    "    \n",
    "    train_loader, test_loader = collateral.get_data_loaders(args)\n",
    "\n",
    "    # optimizer = optim.Adam(model.parameters(), lr=args.lr)\n",
    "    optimizer = optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum)\n",
    "    \n",
    "    test_perfs_char = []\n",
    "    test_perfs_font = []\n",
    "    \n",
    "    for epoch in range(1, args.epochs + args.sabotage_epochs + args.new_adversary_epochs + 1):\n",
    "        initial_phase = epoch <= args.epochs\n",
    "        if initial_phase:\n",
    "            print(\"(initial phase)\")\n",
    "        perturbate = epoch > args.epochs and epoch <= args.epochs + args.sabotage_epochs\n",
    "        if perturbate:\n",
    "            print(\"(perturbate)\")\n",
    "        recover = epoch > args.epochs + args.sabotage_epochs\n",
    "        if recover:\n",
    "            print(\"(recover)\")\n",
    "        collateral.train(\n",
    "            args, model, train_loader, optimizer, epoch, alpha, \n",
    "            initial_phase, perturbate, recover, False\n",
    "        )\n",
    "        test_perf_char, test_perf_font = collateral.test(args, model, test_loader, False)\n",
    "        test_perfs_char.append(test_perf_char)\n",
    "        test_perfs_font.append(test_perf_font)\n",
    "\n",
    "    return test_perfs_char, test_perfs_font"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "We could use the same $Q$ than in part 6, but we could like a better resistance so we rebuild a new one."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 112,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Model loaded!\n"
     ]
    }
   ],
   "source": [
    "path = '../data/models/quadconvnet_part8_out10.pt'\n",
    "path = 'models/test_quadconvnet_0.5_0.002_1.7_10.pt'\n",
    "model = CollateralNet()\n",
    "results = {}\n",
    "\n",
    "try:\n",
    "    model.load_state_dict(torch.load(path))\n",
    "    model.eval()\n",
    "    print('Model loaded!')\n",
    "except FileNotFoundError:\n",
    "    print('Computing model...')\n",
    "    alpha=1.7\n",
    "    test_perfs_char_perturbate, test_perfs_font_perturbate = build_resistance(model, alpha=alpha)\n",
    "    results[f\"Main task with perturbation alpha={alpha}\"] = test_perfs_char_perturbate\n",
    "    results[f\"Collateral task with perturbation alpha={alpha}\"] = test_perfs_font_perturbate\n",
    "    \n",
    "    model.results = results\n",
    "    # save model\n",
    "    torch.save(model.state_dict(), path)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 8.2 Testing distinguishability againts multiples nets"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "We load the `Q` net that we have thoroughly built and it will be given to our new network structure"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 113,
   "metadata": {},
   "outputs": [],
   "source": [
    "quadratic_model = model\n",
    "quadratic_model.freeze('quad')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "We now perform only the last phase with the new adversary, note that we change the data loader"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 122,
   "metadata": {},
   "outputs": [],
   "source": [
    "fonts = ['cursive', 'Georgia']\n",
    "letter = \"6\"\n",
    "\n",
    "def evaluate_resistance(model, alpha=0):\n",
    "    args = Parser()\n",
    "    \n",
    "    train_loader, test_loader = collateral.data.get_collateral_data_loaders(args, *fonts, letter)\n",
    "\n",
    "    optimizer = optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum)\n",
    "    \n",
    "    test_perfs_char = []\n",
    "    test_perfs_font = []\n",
    "    \n",
    "    for epoch in range(1, args.new_adversary_epochs + 1):\n",
    "        \n",
    "        print(model.proj1.weight[30:35, 60])\n",
    "        initial_phase = False\n",
    "        perturbate = False\n",
    "        recover = True\n",
    "        \n",
    "        new_adversary = True\n",
    "        \n",
    "        collateral.train(\n",
    "            args, model, train_loader, optimizer, epoch, alpha,\n",
    "            initial_phase, perturbate, recover, new_adversary\n",
    "        )\n",
    "        test_perf_char, test_perf_font = collateral.test(args, model, test_loader, new_adversary)\n",
    "        test_perfs_char.append(test_perf_char)\n",
    "        test_perfs_font.append(test_perf_font)\n",
    "\n",
    "    return test_perfs_char, test_perfs_font"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Here is the basic structure of a network like those seen previously. There is a quadratic part which inherits from the `Q` loaded, and a method `conv_font` is left empty so that we can provide arbitrary adversary by inheriting from this `BaseNet`."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 123,
   "metadata": {},
   "outputs": [],
   "source": [
    "class BaseNet(nn.Module):\n",
    "    def __init__(self, quadratic_model):\n",
    "        super(BaseNet, self).__init__()\n",
    "        self.proj1 = quadratic_model.proj1\n",
    "        self.diag1 = quadratic_model.diag1\n",
    "        \n",
    "    def quad(self, x):\n",
    "        \"\"\"Same as forward up to the junction part\n",
    "        Used for the collateral training\"\"\"\n",
    "        # --- Quadratic \n",
    "        x = x.view(-1, 784)\n",
    "        x = self.proj1(x)\n",
    "        x = x * x\n",
    "        x = self.diag1(x)\n",
    "        return x\n",
    "    \n",
    "    def conv_font(self, x):\n",
    "        pass\n",
    "    \n",
    "    def forward_adv_font(self, x):\n",
    "        x = self.quad(x)\n",
    "        x = self.conv_font(x)\n",
    "        return F.log_softmax(x, dim=1)\n",
    "    \n",
    "    def get_params(self, net):\n",
    "        \"\"\"Select the params for a given part of the net\"\"\"\n",
    "        if net == 'quad':\n",
    "            layers = [self.proj1, self.diag1]\n",
    "        else:\n",
    "            raise AttributeError(f'{net} type not recognized')\n",
    "        params = [p for layer in layers for p in layer.parameters()]\n",
    "        return params\n",
    "    \n",
    "    def freeze(self, net):\n",
    "        \"\"\"Freeze a part of the net\"\"\"\n",
    "        net_params = self.get_params(net)\n",
    "        for param in net_params:\n",
    "            param.requires_grad = False\n",
    "            \n",
    "    def unfreeze(self):\n",
    "        \"\"\"Unfreeze the net\"\"\"\n",
    "        for param in self.parameters():\n",
    "            param.requires_grad = True"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 124,
   "metadata": {},
   "outputs": [],
   "source": [
    "resistance_reports = {}"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 6.2.1 Fully connected models "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Here we inherit from BaseNet and we only have to specify `conv_font` and declare the layers used."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 125,
   "metadata": {},
   "outputs": [],
   "source": [
    "class FFNet(BaseNet):\n",
    "    def __init__(self, architecture, quadratic_model):\n",
    "        super(FFNet, self).__init__(quadratic_model)\n",
    "        # --- FFNs for font families\n",
    "        self.architecture = architecture\n",
    "        n_layer = len(architecture) + 1\n",
    "        input_size = PRIVATE_OUTPUT_SIZE\n",
    "        for i_layer, output_size in enumerate(architecture):\n",
    "            setattr(self, f\"net_{i_layer}\", nn.Linear(input_size, output_size))\n",
    "            input_size = output_size\n",
    "        setattr(self, f\"net_{n_layer}\", nn.Linear(input_size, N_FONTS)) \n",
    "    \n",
    "    def conv_font(self, x):\n",
    "        # --- FFN\n",
    "        architecture = self.architecture\n",
    "        n_layer = len(architecture) + 1\n",
    "        for i_layer, output_size in enumerate(architecture):\n",
    "            linear = getattr(self, f\"net_{i_layer}\")\n",
    "            x = F.relu(linear(x))\n",
    "        linear = getattr(self, f\"net_{n_layer}\")  \n",
    "        x = linear(x)\n",
    "        return x\n",
    "        \n",
    "        "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 126,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n",
      "----- [64, 32, 16, 8]\n",
      "FFNet(\n",
      "  (proj1): Linear(in_features=784, out_features=40, bias=True)\n",
      "  (diag1): Linear(in_features=40, out_features=10, bias=False)\n",
      "  (net_0): Linear(in_features=10, out_features=64, bias=True)\n",
      "  (net_1): Linear(in_features=64, out_features=32, bias=True)\n",
      "  (net_2): Linear(in_features=32, out_features=16, bias=True)\n",
      "  (net_3): Linear(in_features=16, out_features=8, bias=True)\n",
      "  (net_5): Linear(in_features=8, out_features=2, bias=True)\n",
      ")\n",
      "Training set 60000 items\n",
      "Testing set  10000 items\n",
      "tensor([-0.0020, -0.0108,  0.0148, -0.0320, -0.0066])\n",
      "Train Epoch: 1 [0/60000 (0%)]\tLoss Char: 0.000000 Loss Font: 0.718810\n",
      "Train Epoch: 1 [19200/60000 (32%)]\tLoss Char: 0.000000 Loss Font: 0.684452\n",
      "Train Epoch: 1 [38400/60000 (64%)]\tLoss Char: 0.000000 Loss Font: 0.686363\n",
      "Train Epoch: 1 [57600/60000 (96%)]\tLoss Char: 0.000000 Loss Font: 0.688123\n",
      "\n",
      "Test set: Accuracy Char : 0/10000 (0.00%)\n",
      "          Accuracy Font : 5239/10000 (52.39%)\n",
      "tensor([-0.0020, -0.0108,  0.0148, -0.0320, -0.0066], grad_fn=<SelectBackward>)\n",
      "Train Epoch: 2 [0/60000 (0%)]\tLoss Char: 0.000000 Loss Font: 0.660340\n",
      "Train Epoch: 2 [19200/60000 (32%)]\tLoss Char: 0.000000 Loss Font: 0.645407\n",
      "Train Epoch: 2 [38400/60000 (64%)]\tLoss Char: 0.000000 Loss Font: 0.664727\n",
      "Train Epoch: 2 [57600/60000 (96%)]\tLoss Char: 0.000000 Loss Font: 0.647761\n",
      "\n",
      "Test set: Accuracy Char : 0/10000 (0.00%)\n",
      "          Accuracy Font : 6594/10000 (65.94%)\n",
      "tensor([-0.0020, -0.0108,  0.0148, -0.0320, -0.0066], grad_fn=<SelectBackward>)\n",
      "Train Epoch: 3 [0/60000 (0%)]\tLoss Char: 0.000000 Loss Font: 0.634458\n",
      "Train Epoch: 3 [19200/60000 (32%)]\tLoss Char: 0.000000 Loss Font: 0.587245\n",
      "Train Epoch: 3 [38400/60000 (64%)]\tLoss Char: 0.000000 Loss Font: 0.618701\n",
      "Train Epoch: 3 [57600/60000 (96%)]\tLoss Char: 0.000000 Loss Font: 0.582242\n",
      "\n",
      "Test set: Accuracy Char : 0/10000 (0.00%)\n",
      "          Accuracy Font : 6587/10000 (65.87%)\n",
      "tensor([-0.0020, -0.0108,  0.0148, -0.0320, -0.0066], grad_fn=<SelectBackward>)\n",
      "Train Epoch: 4 [0/60000 (0%)]\tLoss Char: 0.000000 Loss Font: 0.643366\n",
      "Train Epoch: 4 [19200/60000 (32%)]\tLoss Char: 0.000000 Loss Font: 0.533711\n",
      "Train Epoch: 4 [38400/60000 (64%)]\tLoss Char: 0.000000 Loss Font: 0.620570\n",
      "Train Epoch: 4 [57600/60000 (96%)]\tLoss Char: 0.000000 Loss Font: 0.603983\n",
      "\n",
      "Test set: Accuracy Char : 0/10000 (0.00%)\n",
      "          Accuracy Font : 6735/10000 (67.35%)\n",
      "tensor([-0.0020, -0.0108,  0.0148, -0.0320, -0.0066], grad_fn=<SelectBackward>)\n",
      "Train Epoch: 5 [0/60000 (0%)]\tLoss Char: 0.000000 Loss Font: 0.567920\n",
      "Train Epoch: 5 [19200/60000 (32%)]\tLoss Char: 0.000000 Loss Font: 0.613715\n",
      "Train Epoch: 5 [38400/60000 (64%)]\tLoss Char: 0.000000 Loss Font: 0.588555\n",
      "Train Epoch: 5 [57600/60000 (96%)]\tLoss Char: 0.000000 Loss Font: 0.588048\n",
      "\n",
      "Test set: Accuracy Char : 0/10000 (0.00%)\n",
      "          Accuracy Font : 6596/10000 (65.96%)\n",
      "tensor([-0.0020, -0.0108,  0.0148, -0.0320, -0.0066], grad_fn=<SelectBackward>)\n",
      "Train Epoch: 6 [0/60000 (0%)]\tLoss Char: 0.000000 Loss Font: 0.551376\n",
      "Train Epoch: 6 [19200/60000 (32%)]\tLoss Char: 0.000000 Loss Font: 0.560764\n",
      "Train Epoch: 6 [38400/60000 (64%)]\tLoss Char: 0.000000 Loss Font: 0.613330\n",
      "Train Epoch: 6 [57600/60000 (96%)]\tLoss Char: 0.000000 Loss Font: 0.508350\n",
      "\n",
      "Test set: Accuracy Char : 0/10000 (0.00%)\n",
      "          Accuracy Font : 6445/10000 (64.45%)\n",
      "tensor([-0.0020, -0.0108,  0.0148, -0.0320, -0.0066], grad_fn=<SelectBackward>)\n",
      "Train Epoch: 7 [0/60000 (0%)]\tLoss Char: 0.000000 Loss Font: 0.684851\n",
      "Train Epoch: 7 [19200/60000 (32%)]\tLoss Char: 0.000000 Loss Font: 0.579256\n",
      "Train Epoch: 7 [38400/60000 (64%)]\tLoss Char: 0.000000 Loss Font: 0.512023\n",
      "Train Epoch: 7 [57600/60000 (96%)]\tLoss Char: 0.000000 Loss Font: 0.622171\n",
      "\n",
      "Test set: Accuracy Char : 0/10000 (0.00%)\n",
      "          Accuracy Font : 7073/10000 (70.73%)\n",
      "tensor([-0.0020, -0.0108,  0.0148, -0.0320, -0.0066], grad_fn=<SelectBackward>)\n",
      "Train Epoch: 8 [0/60000 (0%)]\tLoss Char: 0.000000 Loss Font: 0.574226\n",
      "Train Epoch: 8 [19200/60000 (32%)]\tLoss Char: 0.000000 Loss Font: 0.544816\n",
      "Train Epoch: 8 [38400/60000 (64%)]\tLoss Char: 0.000000 Loss Font: 0.603654\n",
      "Train Epoch: 8 [57600/60000 (96%)]\tLoss Char: 0.000000 Loss Font: 0.565505\n",
      "\n",
      "Test set: Accuracy Char : 0/10000 (0.00%)\n",
      "          Accuracy Font : 7248/10000 (72.48%)\n",
      "tensor([-0.0020, -0.0108,  0.0148, -0.0320, -0.0066], grad_fn=<SelectBackward>)\n",
      "Train Epoch: 9 [0/60000 (0%)]\tLoss Char: 0.000000 Loss Font: 0.587445\n",
      "Train Epoch: 9 [19200/60000 (32%)]\tLoss Char: 0.000000 Loss Font: 0.567299\n",
      "Train Epoch: 9 [38400/60000 (64%)]\tLoss Char: 0.000000 Loss Font: 0.533782\n",
      "Train Epoch: 9 [57600/60000 (96%)]\tLoss Char: 0.000000 Loss Font: 0.590205\n",
      "\n",
      "Test set: Accuracy Char : 0/10000 (0.00%)\n",
      "          Accuracy Font : 6715/10000 (67.15%)\n",
      "tensor([-0.0020, -0.0108,  0.0148, -0.0320, -0.0066], grad_fn=<SelectBackward>)\n",
      "Train Epoch: 10 [0/60000 (0%)]\tLoss Char: 0.000000 Loss Font: 0.613353\n",
      "Train Epoch: 10 [19200/60000 (32%)]\tLoss Char: 0.000000 Loss Font: 0.500187\n",
      "Train Epoch: 10 [38400/60000 (64%)]\tLoss Char: 0.000000 Loss Font: 0.564345\n",
      "Train Epoch: 10 [57600/60000 (96%)]\tLoss Char: 0.000000 Loss Font: 0.524172\n",
      "\n",
      "Test set: Accuracy Char : 0/10000 (0.00%)\n",
      "          Accuracy Font : 7466/10000 (74.66%)\n",
      "tensor([-0.0020, -0.0108,  0.0148, -0.0320, -0.0066], grad_fn=<SelectBackward>)\n",
      "Train Epoch: 11 [0/60000 (0%)]\tLoss Char: 0.000000 Loss Font: 0.432867\n",
      "Train Epoch: 11 [19200/60000 (32%)]\tLoss Char: 0.000000 Loss Font: 0.472330\n",
      "Train Epoch: 11 [38400/60000 (64%)]\tLoss Char: 0.000000 Loss Font: 0.541815\n",
      "Train Epoch: 11 [57600/60000 (96%)]\tLoss Char: 0.000000 Loss Font: 0.539533\n",
      "\n",
      "Test set: Accuracy Char : 0/10000 (0.00%)\n",
      "          Accuracy Font : 7411/10000 (74.11%)\n",
      "tensor([-0.0020, -0.0108,  0.0148, -0.0320, -0.0066], grad_fn=<SelectBackward>)\n",
      "Train Epoch: 12 [0/60000 (0%)]\tLoss Char: 0.000000 Loss Font: 0.472464\n",
      "Train Epoch: 12 [19200/60000 (32%)]\tLoss Char: 0.000000 Loss Font: 0.510915\n",
      "Train Epoch: 12 [38400/60000 (64%)]\tLoss Char: 0.000000 Loss Font: 0.579518\n",
      "Train Epoch: 12 [57600/60000 (96%)]\tLoss Char: 0.000000 Loss Font: 0.436813\n",
      "\n",
      "Test set: Accuracy Char : 0/10000 (0.00%)\n",
      "          Accuracy Font : 7752/10000 (77.52%)\n",
      "tensor([-0.0020, -0.0108,  0.0148, -0.0320, -0.0066], grad_fn=<SelectBackward>)\n",
      "Train Epoch: 13 [0/60000 (0%)]\tLoss Char: 0.000000 Loss Font: 0.435081\n",
      "Train Epoch: 13 [19200/60000 (32%)]\tLoss Char: 0.000000 Loss Font: 0.545237\n",
      "Train Epoch: 13 [38400/60000 (64%)]\tLoss Char: 0.000000 Loss Font: 0.508475\n",
      "Train Epoch: 13 [57600/60000 (96%)]\tLoss Char: 0.000000 Loss Font: 0.474248\n",
      "\n",
      "Test set: Accuracy Char : 0/10000 (0.00%)\n",
      "          Accuracy Font : 7871/10000 (78.71%)\n",
      "tensor([-0.0020, -0.0108,  0.0148, -0.0320, -0.0066], grad_fn=<SelectBackward>)\n",
      "Train Epoch: 14 [0/60000 (0%)]\tLoss Char: 0.000000 Loss Font: 0.407262\n",
      "Train Epoch: 14 [19200/60000 (32%)]\tLoss Char: 0.000000 Loss Font: 0.444260\n",
      "Train Epoch: 14 [38400/60000 (64%)]\tLoss Char: 0.000000 Loss Font: 0.450477\n"
     ]
    },
    {
     "ename": "KeyboardInterrupt",
     "evalue": "",
     "output_type": "error",
     "traceback": [
      "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
      "\u001b[0;31mKeyboardInterrupt\u001b[0m                         Traceback (most recent call last)",
      "\u001b[0;32m<ipython-input-126-d51ce4c95435>\u001b[0m in \u001b[0;36m<module>\u001b[0;34m\u001b[0m\n\u001b[1;32m      5\u001b[0m     \u001b[0mmodel\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mFFNet\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0marchitecture\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mquadratic_model\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m      6\u001b[0m     \u001b[0mprint\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mmodel\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m----> 7\u001b[0;31m     \u001b[0m_\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mtest_perfs_font\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mevaluate_resistance\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mmodel\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m      8\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m      9\u001b[0m     \u001b[0marchitecture\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0;34m':'\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mjoin\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mmap\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mstr\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m[\u001b[0m\u001b[0;36m8\u001b[0m\u001b[0;34m]\u001b[0m \u001b[0;34m+\u001b[0m \u001b[0marchitecture\u001b[0m \u001b[0;34m+\u001b[0m \u001b[0;34m[\u001b[0m\u001b[0mN_FONTS\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;32m<ipython-input-122-ca5d34b41410>\u001b[0m in \u001b[0;36mevaluate_resistance\u001b[0;34m(model, alpha)\u001b[0m\n\u001b[1;32m     23\u001b[0m         collateral.train(\n\u001b[1;32m     24\u001b[0m             \u001b[0margs\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mmodel\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mtrain_loader\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0moptimizer\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mepoch\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0malpha\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 25\u001b[0;31m             \u001b[0minitial_phase\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mperturbate\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mrecover\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mnew_adversary\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m     26\u001b[0m         )\n\u001b[1;32m     27\u001b[0m         \u001b[0mtest_perf_char\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mtest_perf_font\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mcollateral\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mtest\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0margs\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mmodel\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mtest_loader\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mnew_adversary\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;32m~/code/collateral-learning/learn/collateral/train.py\u001b[0m in \u001b[0;36mtrain\u001b[0;34m(args, model, train_loader, optimizer, epoch, alpha, initial_phase, perturbate, recover, new_collateral)\u001b[0m\n\u001b[1;32m     45\u001b[0m             \u001b[0mloss_font\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mF\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mnll_loss\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0moutput_font\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mtarget_font\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     46\u001b[0m             \u001b[0mloss_font\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mbackward\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 47\u001b[0;31m             \u001b[0moptimizer\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mstep\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m     48\u001b[0m             \u001b[0mmodel\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0munfreeze\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     49\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;32m~/code/env/pysyft/lib/python3.7/site-packages/torch/optim/sgd.py\u001b[0m in \u001b[0;36mstep\u001b[0;34m(self, closure)\u001b[0m\n\u001b[1;32m     93\u001b[0m                     \u001b[0md_p\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0madd_\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mweight_decay\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mp\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mdata\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     94\u001b[0m                 \u001b[0;32mif\u001b[0m \u001b[0mmomentum\u001b[0m \u001b[0;34m!=\u001b[0m \u001b[0;36m0\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 95\u001b[0;31m                     \u001b[0mparam_state\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mstate\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0mp\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m     96\u001b[0m                     \u001b[0;32mif\u001b[0m \u001b[0;34m'momentum_buffer'\u001b[0m \u001b[0;32mnot\u001b[0m \u001b[0;32min\u001b[0m \u001b[0mparam_state\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     97\u001b[0m                         \u001b[0mbuf\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mparam_state\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0;34m'momentum_buffer'\u001b[0m\u001b[0;34m]\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mtorch\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mzeros_like\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mp\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mdata\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;32m~/code/env/pysyft/lib/python3.7/site-packages/torch/tensor.py\u001b[0m in \u001b[0;36m__hash__\u001b[0;34m(self)\u001b[0m\n\u001b[1;32m    428\u001b[0m         \u001b[0;32mreturn\u001b[0m \u001b[0miter\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mimap\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;32mlambda\u001b[0m \u001b[0mi\u001b[0m\u001b[0;34m:\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0mi\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mrange\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0msize\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;36m0\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m    429\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m--> 430\u001b[0;31m     \u001b[0;32mdef\u001b[0m \u001b[0m__hash__\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m    431\u001b[0m         \u001b[0;32mreturn\u001b[0m \u001b[0mid\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m    432\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;31mKeyboardInterrupt\u001b[0m: "
     ]
    }
   ],
   "source": [
    "architectures = [[64, 32, 16, 8], [32, 16, 8], [24, 12], [64], [32], [16]]\n",
    "\n",
    "for architecture in architectures:\n",
    "    print('\\n-----', architecture)\n",
    "    model = FFNet(architecture, quadratic_model)\n",
    "    print(model)\n",
    "    _, test_perfs_font = evaluate_resistance(model)\n",
    "    \n",
    "    architecture = ':'.join(map(str, [8] + architecture + [N_FONTS]))\n",
    "    resistance_reports[f\"Collateral task with net {architecture}\"] = test_perfs_font\n",
    "    \n",
    "show_results(resistance_reports, title=\"Resistance of FFNs with CNN protection\")\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 6.2.2 CNN models"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [],
   "source": [
    "class CNNet(BaseNet):\n",
    "    def __init__(self, nn_modules, quadratic_model):\n",
    "        super(CNNet, self).__init__(quadratic_model)\n",
    "        # --- CNNs for font families\n",
    "        self.nn_modules = nn_modules\n",
    "        setattr(self, \"net_0\", nn.Linear(PRIVATE_OUTPUT_SIZE, 784))\n",
    "        for i_layer, nn_module in enumerate(nn_modules):\n",
    "            setattr(self, f\"net_{i_layer + 1}\", nn_module)\n",
    "    \n",
    "    def conv_font(self, x):\n",
    "        switched_from_conv_lin = False\n",
    "        out_channels = []\n",
    "        # Make the junction\n",
    "        linear = getattr(self, \"net_0\")  \n",
    "        x = linear(x)\n",
    "        x = x.view(-1, 1, 28, 28)\n",
    "        # --- CNN\n",
    "        for i_layer, _ in enumerate(self.nn_modules):\n",
    "            layer = getattr(self, f\"net_{i_layer + 1}\")\n",
    "            if isinstance(layer, nn.Conv2d): # Conv layer\n",
    "                x = F.relu(layer(x))\n",
    "                x = F.max_pool2d(x, 2, 2)\n",
    "                out_channels.append(layer.out_channels)\n",
    "            else: # Linear layer\n",
    "                if not switched_from_conv_lin:\n",
    "                    x = x.view(-1, 4*4*out_channels[-1])\n",
    "                    switched_from_conv_lin = True\n",
    "                if i_layer < len(self.nn_modules) - 1:\n",
    "                    x = F.relu(layer(x))\n",
    "                else:\n",
    "                    x = layer(x)\n",
    "        return x\n",
    "         "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Training set 60000 items\n",
      "Testing set  10000 items\n",
      "Train Epoch: 1 [0/60000 (0%)]\tLoss Char: 0.000000 Loss Font: 0.911390\n",
      "Train Epoch: 1 [19200/60000 (32%)]\tLoss Char: 0.000000 Loss Font: 0.667801\n",
      "Train Epoch: 1 [38400/60000 (64%)]\tLoss Char: 0.000000 Loss Font: 0.685084\n",
      "Train Epoch: 1 [57600/60000 (96%)]\tLoss Char: 0.000000 Loss Font: 0.683268\n",
      "\n",
      "Test set: Accuracy Char : 0/10000 (0.00%)\n",
      "          Accuracy Font : 5860/10000 (58.60%)\n",
      "Train Epoch: 2 [0/60000 (0%)]\tLoss Char: 0.000000 Loss Font: 0.650930\n",
      "Train Epoch: 2 [19200/60000 (32%)]\tLoss Char: 0.000000 Loss Font: 0.662303\n",
      "Train Epoch: 2 [38400/60000 (64%)]\tLoss Char: 0.000000 Loss Font: 0.687251\n",
      "Train Epoch: 2 [57600/60000 (96%)]\tLoss Char: 0.000000 Loss Font: 0.638498\n",
      "\n",
      "Test set: Accuracy Char : 0/10000 (0.00%)\n",
      "          Accuracy Font : 6114/10000 (61.14%)\n",
      "Train Epoch: 3 [0/60000 (0%)]\tLoss Char: 0.000000 Loss Font: 0.664561\n",
      "Train Epoch: 3 [19200/60000 (32%)]\tLoss Char: 0.000000 Loss Font: 0.618946\n",
      "Train Epoch: 3 [38400/60000 (64%)]\tLoss Char: 0.000000 Loss Font: 0.624185\n",
      "Train Epoch: 3 [57600/60000 (96%)]\tLoss Char: 0.000000 Loss Font: 0.675070\n",
      "\n",
      "Test set: Accuracy Char : 0/10000 (0.00%)\n",
      "          Accuracy Font : 6415/10000 (64.15%)\n",
      "Train Epoch: 4 [0/60000 (0%)]\tLoss Char: 0.000000 Loss Font: 0.626867\n",
      "Train Epoch: 4 [19200/60000 (32%)]\tLoss Char: 0.000000 Loss Font: 0.614342\n",
      "Train Epoch: 4 [38400/60000 (64%)]\tLoss Char: 0.000000 Loss Font: 0.562274\n",
      "Train Epoch: 4 [57600/60000 (96%)]\tLoss Char: 0.000000 Loss Font: 0.590366\n",
      "\n",
      "Test set: Accuracy Char : 0/10000 (0.00%)\n",
      "          Accuracy Font : 6544/10000 (65.44%)\n",
      "Train Epoch: 5 [0/60000 (0%)]\tLoss Char: 0.000000 Loss Font: 0.596705\n",
      "Train Epoch: 5 [19200/60000 (32%)]\tLoss Char: 0.000000 Loss Font: 0.628992\n",
      "Train Epoch: 5 [38400/60000 (64%)]\tLoss Char: 0.000000 Loss Font: 0.642429\n",
      "Train Epoch: 5 [57600/60000 (96%)]\tLoss Char: 0.000000 Loss Font: 0.564583\n",
      "\n",
      "Test set: Accuracy Char : 0/10000 (0.00%)\n",
      "          Accuracy Font : 6614/10000 (66.14%)\n",
      "Train Epoch: 6 [0/60000 (0%)]\tLoss Char: 0.000000 Loss Font: 0.592426\n",
      "Train Epoch: 6 [19200/60000 (32%)]\tLoss Char: 0.000000 Loss Font: 0.574638\n",
      "Train Epoch: 6 [38400/60000 (64%)]\tLoss Char: 0.000000 Loss Font: 0.579814\n",
      "Train Epoch: 6 [57600/60000 (96%)]\tLoss Char: 0.000000 Loss Font: 0.559498\n",
      "\n",
      "Test set: Accuracy Char : 0/10000 (0.00%)\n",
      "          Accuracy Font : 6295/10000 (62.95%)\n",
      "Train Epoch: 7 [0/60000 (0%)]\tLoss Char: 0.000000 Loss Font: 0.662299\n",
      "Train Epoch: 7 [19200/60000 (32%)]\tLoss Char: 0.000000 Loss Font: 0.551563\n",
      "Train Epoch: 7 [38400/60000 (64%)]\tLoss Char: 0.000000 Loss Font: 0.585619\n",
      "Train Epoch: 7 [57600/60000 (96%)]\tLoss Char: 0.000000 Loss Font: 0.636200\n",
      "\n",
      "Test set: Accuracy Char : 0/10000 (0.00%)\n",
      "          Accuracy Font : 6878/10000 (68.78%)\n",
      "Train Epoch: 8 [0/60000 (0%)]\tLoss Char: 0.000000 Loss Font: 0.652154\n",
      "Train Epoch: 8 [19200/60000 (32%)]\tLoss Char: 0.000000 Loss Font: 0.625389\n",
      "Train Epoch: 8 [38400/60000 (64%)]\tLoss Char: 0.000000 Loss Font: 0.597918\n",
      "Train Epoch: 8 [57600/60000 (96%)]\tLoss Char: 0.000000 Loss Font: 0.614232\n",
      "\n",
      "Test set: Accuracy Char : 0/10000 (0.00%)\n",
      "          Accuracy Font : 7013/10000 (70.13%)\n",
      "Train Epoch: 9 [0/60000 (0%)]\tLoss Char: 0.000000 Loss Font: 0.620657\n",
      "Train Epoch: 9 [19200/60000 (32%)]\tLoss Char: 0.000000 Loss Font: 0.599668\n",
      "Train Epoch: 9 [38400/60000 (64%)]\tLoss Char: 0.000000 Loss Font: 0.743383\n",
      "Train Epoch: 9 [57600/60000 (96%)]\tLoss Char: 0.000000 Loss Font: 0.546876\n",
      "\n",
      "Test set: Accuracy Char : 0/10000 (0.00%)\n",
      "          Accuracy Font : 6930/10000 (69.30%)\n",
      "Train Epoch: 10 [0/60000 (0%)]\tLoss Char: 0.000000 Loss Font: 0.595817\n",
      "Train Epoch: 10 [19200/60000 (32%)]\tLoss Char: 0.000000 Loss Font: 0.587947\n",
      "Train Epoch: 10 [38400/60000 (64%)]\tLoss Char: 0.000000 Loss Font: 0.594132\n",
      "Train Epoch: 10 [57600/60000 (96%)]\tLoss Char: 0.000000 Loss Font: 0.577100\n",
      "\n",
      "Test set: Accuracy Char : 0/10000 (0.00%)\n",
      "          Accuracy Font : 7090/10000 (70.90%)\n"
     ]
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAmsAAAEWCAYAAAA5GNBmAAAABHNCSVQICAgIfAhkiAAAAAlwSFlzAAALEgAACxIB0t1+/AAAADl0RVh0U29mdHdhcmUAbWF0cGxvdGxpYiB2ZXJzaW9uIDMuMC4zLCBodHRwOi8vbWF0cGxvdGxpYi5vcmcvnQurowAAIABJREFUeJzsnXd4VFX+h98zJTNpM5l0AoQk9I6AigV7dxXEioqoq1hWd61r33V3dfW3YnddwK7oWlYEdVkrglhQAQHphCS09DaZZPrc8/vj3oQEEhIgIQOe93nyzMy955x77p07mc982xFSShQKhUKhUCgU0YmpuyegUCgUCoVCoWgbJdYUCoVCoVAoohgl1hQKhUKhUCiiGCXWFAqFQqFQKKIYJdYUCoVCoVAoohgl1hQKhUKhUCiiGCXWFIooRgixRghxQnfPoysQQtwghCgTQtQLIVK6ez6dRXvvmRBioRDimgM4pUMe4x7K6+55KBRdhRJrCkUnIIQoEkL4jC+NUiHEq0KIhP0dV0o5VEq5sJ1j5wghpBDCsr/HO1AIIazAE8BpUsoEKWVVK21ihBAPCiE2CSEajGv8shAix9i/UAjhF0L0btbnFCFEUbPXRUKIciFEfLNt1wghFnbVuTV/z4z5z96f8YQQDiHEU0KIrcb9tdl4nWrsb/ccjfvjFyGEqdm2h4QQr+7P3PbiHK4UQnzTSWPtJnaNe6igM8ZXKKIRJdYUis7jHCllAjAKOAy4p5vnE81kAHZgzR7a/Ac4F7gUcAIjgWXAyc3aNAAPtHMsM/CHfZ5pNyKEiAG+BIYCZwAO4CigCjiiWdOOnGMWcEkXTBMhhLkrxlUoFDpKrCkUnYyUshT4FF20ASCEsAkhphvWkTIhxAwhRKyxL1UI8bEQolYIUS2EWNxoATGsJqcYz48QQiwVQtQZYzxhDP+18VhrWF6OEkL0FUIsEEJUCSEqhRBvCiGSms2nSAhxhxBilRDCLYR4Rwhhb7Z/ghBihXGszUKIM4ztTiHES0KIEiHEDsM60+oXtXHOTwkhio2/p4xtA4ANzea8oJW+pwCnAhOklD9JKcNSSreU8p9SypeaNX0GmCyE6LuHt+Qx4I7m59/sOEII8aRhmaozrE/DWml3ohDil2avPxdC/NTs9WIhxMRm1/YU45rdC1xsvC8rmw3ZRwjxrRDCI4T4rNFK1gpXANnAeVLKtVJKTUpZLqX8m5RyfkfOsRn/AP7SEQusEOIEIcR2IcS9xv1TJIS4rNn+V4UQ/xJCzBdCNAAnGvfG60KICiHEFiHE/UIIkxBiMDADOMq4DrXGGG1+Joz9u92DQoiHgfHAc8ZYzxltpRCin/G81XkY+64UQnxjHLdGCFEohDizveuhUHQ3SqwpFJ2MEKIXcCaQ32zzo8AAdAHXD+gJ/MnYdzuwHUhDtzjdC7S2DtzTwNNSSgfQF3jX2H6c8ZhkuIO+BwTwCLo1ZTDQG3hwl/EuQrfW5AIjgCuN+R8BvA7cCSQZ4xcZfV4FwsY5HAacBrQVf3UfMM4455HolqD7pZQb0S1FjXM+qZW+pwA/Sim3tTF2IzuAF4C/7KHNUmAhcEcr+05DP78B6Na7i9CtVruyBOgvdGFtRb9eWUKIRENgjAUWN+8gpfwE+DvwjvG+jGy2+1LgKiAdiGljbqBfh0+klPV7OL/2zrGROUAdxvvcATKBVPR7dSowSwgxsNn+S4GHgUTgG+BZ9GuYBxyPLjSvklKuA64HvjeuQ6OgbPMz0dY9KKW8D/0632SMdVMr8251Hs32H4n+YyEVXcC+JIQQHbwmCkW3oMSaQtF5zBVCeIBtQDnwZ9CtN8A04FYpZbWU0oP+Jd7okgoBPYA+UsqQlHKxbH3R3hDQTwiRKqWsl1IuaWsiUsp8KeXnUsqAlLICPT7s+F2aPSOlLJZSVgMfsdMS+FvgZaO/JqXcIaVcL4TIAM4CbpFSNkgpy4Enadu1dhnwV8MSVIEuqKa0NeddSAFKOtj2EeAcIcTQPbT5E3CzECJtl+0hdLExCBBSynVSyt2OK6X0AT+hi4YxwErgW+AYdEG6qbW4uz3wipRyozHuuzSzwu7C3lyHts6xEYnuMn5A6O7VjvCAcQ8tAv6LLmYbmSel/FZKqaFfx0uAe6SUHillEfA4bbzfHfhMtHoPtjdZw8rb3jy2SClfkFJGgNfQP3sZHbscCkX3oMSaQtF5TJRSJgInoH/5N7q20oA4YJnQXZ21wCfGdtBdWPnAZ0KIAiHE3W2M/1t0S8R6IcRPQojftDURIUSGEOJtw1VZB8xuNp9GSps99wKNCRG9gc2tDNsHsAIlzc5jJrp1qDWygC3NXm8xtnWEKvQv0XYxhOBzwF/30GY18DFw9y7bFxh9/wmUCyFmCSEcbQyzCP29Pc54vhBdAB9vvN4b2rr2u7I316HVc9ylzXx0K+51HRiyRkrZ0Oz1ru9fc6tnKvq9sev73bONsdv7TLR1D7ZHR+bRdO2llF7j6X4nAykUXYkSawpFJ2NYIV4FphubKgEfMFRKmWT8OY1kBAwLwO1Syjz0gPrbhBAntzLuJinlZHRx9H/Af4SeAdiaFe7vxvbhhtv0cnTXaEfYhu5mbW17AEhtdh4OKWVbFq1idIHXSLaxrSN8ARxhuJQ7wmPAiehWr7b4M3AtuwgIKeUzUsoxwBB0MXxnG/13FWuLaF+stfbe7A1fAKeLZpme7dDqOe7Cfeiu9rh2xnLtctxd37/m51aJbl3b9f3e0UrbxvZtfiZo+x5sbaxdx93TPBSKgxIl1hSKruEp4FQhxEjDTfQC8KQQIh1ACNFTCHG68fw3Qoh+hmvIDUQAbdcBhRCXCyHSjPFqjc0aUGE8Nq8zlQjUA24hRE/aFiCt8RJwlRDiZCNAvKcQYpDhHvwMeFzo5SRMQk9k2NW92si/gfuFEGlGAP2f0C187SKl/AL4HPhACDFGCGEx4sOuF0Jc3Ur7WnR31x/3MGY+8A7w+8ZtQojDhRBHGnFoDYCfVq69wXfAQPTYux+llGvQRcGR7Ezy2JUyIEc0K5mxl7yBLlzeF0IMMq55ihH4f9aujVs7x1baLARWo8ehtcdfhF5CZTzwG+C9NsaMoLtzHzbepz7Abex8v8uAXo3u1/Y+E7RxDzYbq9Waah2Yh0JxUKLEmkLRBRiuudfZmURwF7qrc4nhlvwC/YsfoL/xuh74HnheSvlVK8OeAawRQtSjJxtcIqX0Ga6ch4FvDZfSOPT4sNHo4u+/6MHlHZ37j+gB2U8a/Rex01JxBXpA/FqgBr28RltuuofQA99XAb8Ay41tHeUCYD66+HCjC4yx6NeqNZ5GF7p74q9Ac2uRA1001KC7y6rQrXS7YbgElwNrpJRBY/P36DFQ5W0cr1HcVAkhlrczt9aOGUBPMliPLl7rgB/R3X0/tNFt13NsjfuB5HbalKJfl2LgTeD6duLGbkYXvAXoCQdvAS8b+xagl2kpFUJUGtva/Ey0cw8+DVxgZHM+s5fzUCgOSkTrccwKhUKh+LUi9BUYZkspO+qGVigUXYiyrCkUCoVCoVBEMUqsKRQKhUKhUEQxyg2qUCgUCoVCEcUoy5pCoVAoFApFFNPuGnEHE6mpqTInJ6e7p6FQKBQKheJgQYIWCCN9ETR/GKnpHseYrISOV6fcR5YtW1YppWxr1ZEmDimxlpOTw9KlS7t7GgqFQqFog1BZOeHKCmKH7ml1MIWia4nUBfGtr8K/thp/fi2ENYTNjH2gi9jBKdgHujDFWbt8HkKILe23OsTEmkKhUCiiF83rZevUqQSLiki+8krSbrsVU0xHlylVKPYdKSWhkgb866rxrasitL0eALPLRsIRmdgHJ2PLdSIs0RkdpsSaQqFQKA4IZY88SnDLFhJPP53qV1+l4Ycf6Dn9MWx921pZSqHYd2RYI1DgxreuCv+6aiK1ARAQ0zsRx+l9iB2cgiUjDn3xmOhGiTWFQqFQdDl1n35G7XvvkTJtGum33Yrnq68oue9+CiedT8bdd5F0ySUHxZemIrqJNITwr6/Gv64K/8ZaZDCCsJqw9XfhODkb+6BkzIkHnzX3kCrdMXbsWKli1hQKhSK6CJWUUDDxPGKys8l5602EVY8FCldUUHzvfTQsXkzCiSfS4+GHsCS3twqWQrETKSXhCh/+dVX41lUT3FIHEkyOGGIHJ2MfnIK9rxNhNXf3VFtFCLFMSjm23XZKrCkUCoWiq5CRCFuvuhrf6tXkfTCHmD59Wu7XNGpmv0n59OmYnA6yHnmUhGOP6abZKg4GZEQSKHLjX6db0MJVfgCsWfHYB6cQOzgZa1YCwhT9ltqOirUuc4MKIQaiL8DcSB76otavG9tzgCLgIillTSv9p6IvNgzwkJTyta6aq0KhUCi6hqoXX8L744/0eOSR3YQagDCZSL5iCnFHHknxHbez7ZprSJ56BWm33YbJZuuGGe8fMiKJ1PgxOWIwxUSnNedgRPOH8W+o0ePPNtQgfWEwC2x9k0gY3xP7oBQsSZ1zv1T5qlhfvZ5jekbPj4YDYlkTQpiBHcCRwO+Aainlo0KIuwGXlPKuXdonA0uBsYAElgFjWhN1zVGWNYVCoYgefKtWUXTpZThOO5Wsxx9vNyZN8/spn/44NbNnYxs4UE8+6N//AM1235CanmUYKKglsNlNoNCNDERAgDnJhiUtDmtaLJb0OKxpcVjSYzHFW1V8XgcIV/nwravGv76aQIEbNIkp3oJ9kG49s/VPwmTrXJuTN+Tlt5/+lgJ3AZ+e/ylJ9qROHX9XosoNKoQ4DfizlPIYIcQG4AQpZYkQogewUEo5cJf2k4021xmvZxrt/r2n4yixplAoFNFBpL6BwkmTkOEQeXPnYnY4Oty3ftEiiu+9D62+nvQ/3onr0kujRtxIKQmXewlsduPfXEuw0I3mDQNgSY3F1tdJTM9EIp4goQov4Qof4XIvMqQ1jSFiLbqAS4vDmh6HxRBzFpcdYY6O8+wOpCYJbvM0ldcIl3kBsKTH6fFnQ1KI6Z3YZe7NsBbm9wt+z7fF3/LUCU9xYvaJXXKc5nS7G3QXLgEahVaGlLLEeF4KZLTSviewrdnr7ca23RBCTAOmAWRnZ3fKZBUKhUKxf5Q99BCh7dvp88breyXUABKOP568eXMpvvdeyv72EA1fL6bH3x/GkpLSRbNtGykl4So/gc21+l+BG60+BOg1uuyDU7D1S8KW58TibN0NJzVJpC5AuNynC7hyXcT5N1bjXVa2s6FZYEmJxZquCzndGheLJS220y1I0YIWjBDYVINvbTX+DdX6tTWBLcdJ/G/yiB2cjCUltsvnIaXkoSUPsXjHYh4Y98ABEWp7Q5e/+0KIGOBc4J5d90kppRBiv0x7UspZwCzQLWv7M5ZCoVAo9h/3f/+Le+5cUm+8kbgxY/ZpDEtqKr1nztSTDx57jIIJE8l65O8kjB/fybPdnXC1f6dbc3MtkbogoGcY2vu7sOU5sfVNwpJs79B4wiSwJNmxJNmxD3C12Kf5woaA8xGu8BKq8BEq8+JbWwU7jXGYnTG6gEuLbbLGWdPjMCXGRI3VsaNE3AHdvbmuCv/mWghLhN2MfWCybkEbcGBWD2jOjJUzeH/T+0wbMY2LBl50QI/dEQ6EVD8TWC6lbPz5UCaE6NHMDVreSp8dwAnNXvcCFnbpLBUKhUKx34R27KD0wb8QO2oUqTfesF9jCSFInnI5cUceQfHtd7Dt2mm4pkwh/Y7bOzX5IOIO4C9wN1nPIjUBAEzxVmx9ndjykrD1dWJJje10YWSKtWDLdmDLbml9lGGNcLWfcLku4BqFnHd5uR4TZyBsZl24GfFwVsMiZ0m2R001fikloeKGpvIaoR3G6gHJdhKO7IF9SAq2HAfC3D3zfX/j+zy/8nkm9J3ATaNu6pY5tEeXx6wJId4GPpVSvmK8fgyoapZgkCyl/OMufZLRkwpGG5uWoycYVO/pWCpmTaFQKLoPGQ6z5YqpBDZsIHfeXGJ69eq0sbVAQE8+eOMNbP37k/X4dOwDBuzTWJH6IIEmceYmXOkD9FgyW54Te6PlLAqr20sp0TxBQoYlLlzhI1TuJVzhJeIO7mxoAktybDNL3E4xZ4rtejuNDGn4C2rxr63Cv75an5uAmGwH9sG6Bc2S3v3X9+vtX/P7Bb9nXNY4nj3pWaymA2vRi4oEAyFEPLAVyJNSuo1tKcC7QDawBb10R7UQYixwvZTyGqPd1cC9xlAPN4q9PaHEmkKhUHQfFf/8J5XPPkfWY4/hPOc3XXKM+q+/1pMP6upIv/NOXJdf1n6WqTeki7MCPSmgMXBd2MzYcp1Nbk1rj/iDojZXW2iBsJ7Q0EzAhSp8uhiN7PyuNyVadQvcLkkOZqdtv84/Uh/Ev74a37pqAptqkEENEWPC3t+lF6cd5MKcED2rB/xS8Qu//ey35DpzeeX0V4izxh3wOUSFWDvQKLGmUCgU3YN3+c9smTIFx9ln0fMf/+jSY4Wrqii59z7qFy0i/rjxZP3971hSU5v2a/4wgaK6JrdmqKQBJAiriZgcB7a+ekJATM/EX0X2pYxIwjX+psSGUDOLnPSFm9oJq2mngGssN5IehyUlFmHd3UXZmBnrW1eNf20VwW0ekGB2xGAfYpTXyEtqtW93s7VuK5fPv5w4axyzz5pNamxq+526ACXWFAqFQnFAiHg8FE48D4Qgd+4HmBMSuvyYUkpq3nqL8n88hsnhIu2WBzHFZxPY7Ca4w6MH51uEHg/WV485i+mVGDVxXNGAlBKtIbQzS7W5S7U2oFc5Bb1mnMveVG7EkhbbJNIi1cbqAT0TmpZ3smbFd7t7c09U+iqZMn8KDaEG3jjrDfo4di/WfKCIttIdCoVCoThEKf3LXwmVlpLz5uwDI9RCGoGtdZjTj8V1ZV+CpX4afjIBW7H2dpB4Qm89KaBPYtSuCRkNCCEwJ8RgTojBludssU8LRghXtnSphit8+De7IayBRWDv5yLx+F7EDkrG3EbZkmjDG/Jy05c3Uemr5KXTX2pVqAW3b8e3bBnOCRO6YYato8SaQqFQKPYZ94cfUvfxx6T94ffEjhrVJceQEY3g9vqdtc62eHTBIHSLTuKxqfh+/oLad2Zgy8vGecp07P26tvL8oY4pxkxMVgIxWS3Ft9QkEXcAU7z1oFtOK6SFuH3R7ayrXsczJz7DiLQRLfYH8vOpeuEF3B//F2GzkXDSSZgTE7tpti1RYk2hUCgU+0Rw2zZK//JXYseOIWXatE4bV2qS0I56AgW1+De7CRa5kUG96Ji1RzwJ43roSQF5Tkx2/Wss6Zx+JJ44jOJ77qHowgtJv+N2XFOmRLU77mBEmAQWV8fqy0UTUkr+9v3f+GbHN/z5qD9zfO/jm/b5Vq+hauZMPF98gbDbSZ4yheSrrowaoQZKrCkUCoViH5ChEDvuuANMJnr+4x8I875bWaQmCZU27CynUehG+vVaYpb0OOLGZGDvm0RMrhNzfNulFRLGH0veh/Moufc+yv7+CPWLvyHr7w9jSUvb57kpDg2eX/k8H+R/wPUjr+eCARcA4P3pJypnzqLhm28wORyk3nA9rilTsLhc7Yx24FFiTaFQKBR7TcXzz+NfuYqeTz6BNStrr/pKKQlX+Fou4dS4vmaKnbgRaU3FaM2Je1fqwZKcTK9/PU/t229T9uj/UTBhIj0efojEE6Nr+SDFgeO9je8xY+UMzut3HjeMuIH6r7+mcuYsfMuWYU5JIe3223BNnnxA4i33FSXWFAqFQrFXeH/6iaoZM3FOmoTjzDP3qm/9d8XUfbUVzWOsr5lkrK/Z1yhE2wmB6kIIXJMnE3f44ey4406233AjrksvJf2Pd2KyH3wuPMW+s3DbQh5a8hDjexzLrZ5xbLngQvxr12Lp0YOM++8n6fxJmGK7fu3R/UWV7lAoFApFh4m43RRMPA9TTAy5c97HFB/f4b4NS0up+c8mbH2dxI1Kx5bnxJxs79K4Mi0YpOKJJ6l+9VVi+vWl5/Tp2AcN6rLjKaKHlRUruW7+bzm3MJkLfzATLiwipk8fUqZdi/OccxAx3V+gV5XuUCgUCkWnIqWk5M8PEq6oIOffb+2VUPOtraJmziZsA1ykXjHkgNU7M8XEkHH3XcQfeyzF99xN0YUXkXb7bSRfcQXCpGquHaoUlm/kvf+7hunfBUip2YZ54EAynnicxNNP36/4yu5C3akKhUKh6BDuOR/g+eQTvUzH8OEd7hfYUkfVW+uxZiWQctngbilMm3DsMeR9+CHx48dT/uj/se3aaYTKyw/4PBRdS6S+ga0zn6X8rElc9t960nr2p9e/nid37gc4zjrroBRqoNygCoVCoegAgcJCCs+/gNgRI8h++aUOW6VCZQ2Uz1iFOd5K2vUjun1tSCklte+8S9mjj2KKjdWTD046qVvnpNh/IrW1VM9+k+rXX0erq2N1rpnBt9zH0NMuieryLR11gyrLmkKhUCj2iAwGKb7jTkxWK1n/92iHhVq4NkDly6sRFhOpVw/rdqEGRvLBJReT+/5/sGRmsv3G31Hy4INoPl93T02xD4QrKih77DHyTzqZyueeY1MfK/dPtZLxwvMMO31yVAu1vUHFrCkUCoVij1Q8+yz+NWvo+ewzWDMyOtQn0hCi8uVf0AIR0q4biSU5urIwbX37kvPO21Q89TTVL7+M96el9Jz+GPbBg7t7aooOENqxg6qXXqL2P+8jw2EcZ57JW4f7eT2wiL8c/ReO63Vcd0+xU1GWNYVCoVC0ScOSJVS9+BJJF12E49RTO9RHC0aoem0N4Wo/qVcMJaZHxxMRDiSmmBgy/ngn2S+/hFZXR+FFF1P18itITevuqSnaIFBQQPHd95B/+hnUvPcfnBPOpe//5jPn8j68HljEjSNvZFL/Sd09zU5HWdYUCoVC0SrhmhqK/3gXMbm5ZNx9V4f6yIhG9VvrCW7zkHL54N0WCI9G4o8+mtwP51Fy/wOU/+MfNHyzmB6PPIo1I727p6Yw8K9dS+XMWXg++wxhs+G6dDIpV1+NNTOTdze8y6xVszi///lcP/L67p5ql6DEmkKhUCh2Q0pJyQMPEK6pIXfGvzDFxXWoT837m/CvrybpvH7EDk09ADPtHCwuF72ee5bad9+j7JFHKJwwQU8+OPnk7p7arxrv8uVUzphBw9eLMSUkkDJtGslTr8CSnAzAgq0LePiHhzmu13HcP+7+QyZGbVeUWFMoFArFbtS+8y71X3xJ+l13YR8ypEN93J8U4V1ejuPUPiQc2aOLZ9j5CCFwXXwRcYePpfiOO9n+u5tIuvhiMu6+66Cocn+oIKWk4dvvqJo5E+9PP2F2uUi75RZcl13aYnH1FeUr+OPXf2RoylAeO+4xLKZDV9J06ZkJIZKAF4FhgASuBm4BBhpNkoBaKeWoVvoWAR4gAoQ7ktqqUCgUiv0nsHkzZY8+Svwxx5A89YoO9fEs3k79ou3Ej+tB4km9u3iGXYstL4+ct/9NxTPPUPXSy3h//JGej0/vsGhV7BtS0/B8+SVVM2fhX70aS0YGGffeQ9IFF+xm2S10F3LzgpvJiMvguZOfI87avuX3YKZL66wJIV4DFkspXxRCxABxUsraZvsfB9xSyr+20rcIGCulrOzo8VSdNYVCodg/tGCQoosuJlxWRu68uVjT24/b8v5cTvU7G4gdnkry5EEI06HjimpYsoTiP95FuKaG9FtuIfmqK9XKB52MDIepmz+fylmzCOZvxpqdTcq11+CcMAFTK0tCVfoquXz+5fjCPmafOZvejoP3x0G3LzclhHACxwFXAkgpg0Cw2X4BXASoaoQKhUIRJVQ88SSB9evp9a/nOyTU/BtrqH5vI7a+TpIvHnhICTWA+HHjyJ03l9I//Ynyxx6j/pvFZD36aIdLmCjaRgsGcc/5gKoXXyS0fTu2/v3Jmj4dxxmnIyyty5OGUAM3fnEj1f5qXjn9lYNaqO0NXfnzIBeoAF4RQvwshHhRCNE8f3s8UCal3NRGfwl8JoRYJoSY1oXzVCgUCgVQ/823VL/6Kq5LLyXxxBPbbR/c5qFq9lqsGXGkTDlw630eaCwuFz2feYbMv/0V34qVFJ47gbrPP+/uaR20aF4vVa+8yuZTTqX0wQcxJyfT6/l/kjtvLs7fnN2mUAtFQty28DY21mzk8eMfZ2jq0AM88+6jy9ygQoixwBLgGCnlD0KIp4E6KeUDxv5/AflSysfb6N9TSrlDCJEOfA7cLKX8upV204BpANnZ2WO2bNnSJeejUCgUhzLhqioKJkzE4koi5733MNn3XMQ2VOGlYsZKhM1C+g0jMSe2vzqBpkmKN9YQ8IbpNTgZW+zBFxAeKCyk+I478a9ZQ9KFF5Jxz90dypRVQKSujpo336T6tdeJ1NYSd+SRpF5/HXHjxrWbxSml5P5v7+fDzR/y16P/ynn9zztAs+5aut0NCmwHtkspfzBe/we425icBZgEjGmrs5Ryh/FYLoT4ADgC2E2sSSlnAbNAj1nrzBNQKBSKXwNSSkruvQ+tro6sl15qV6hF6gJUvrQahCDt6mF7FGpSk5QUuMlfWk7+8nJ8dXo0jMks6DXIRe7INHJHphLvtHXqOXUVttxccv79FhXPPkvViy/h/eknsqZPJ3bYr8fKs7eEKyupfu11at56C62hgYQTTiDlumnEHXZYh8d49udn+XDzh/xu1O8OGaG2N3SZWJNSlgohtgkhBkopNwAnA2uN3acA66WU21vra7hLTVJKj/H8NGC3JASFQqFQ7D81b75F/aJFZNx3H/aBA/bYVvOFqXx5NZo3TNp1I7Ck7l7SQkpJ+RYP+UvLyF9WTn1NALPVRM7wFPqNySDOEUPhqkoKVlSw6K0NLPr3BjJzHeSOTCNvVBpJGdFtqRIxMaTffjvxxxxL8V13UTR5Mul/+D3JV1+tkg+aESoupurlV6h97z1kMIjjzDNImTYN+6BBezXO2+vf5oVfXuCCARdw3Yjrumi20U1XZ4OOQi/dEQMUAFdJKWuEEK8CS6SUM5q1zQJelFKeJYTIAz4wdlmAt6SUD7d3PJUNqlAoFHuHf+NGii64kLijxtF7xow9uqNkKELFS6sJbvOQetVQ7P1cO/fp0NVaAAAgAElEQVRJSdWOBvKXlrFpaRl1lX5MZkH20BT6jUknd2QqMfaW9gEpJdXFDRSurKBgRSUVWz0AuHrEkzcqlbxRaaRlJ0Z1odNIbS0lf/ozns8+I+7II8n6v0exZmZ297S6lWBREZUvvIB73ocAOCecS8o112DLzd3rsb7c8iW3LryV43sdz5MnPnnI1VLrqBu0S8XagUaJNYVCoeg4mt9P0YUXEa6pIW/eXCwpKW22lZqkavY6/OuqSJ48iLgRaQDUlDaQv6ycTT+VUVPqRZh092a/MenkjUrDHm/t8Hw81X4KVlRQuLKC4k1upCZJcNl0V+moVLL6J2E2R5/lSkqJe84cSh/+O8JqJfOBB4gbMxpTXBym+Pg2A+YPNfwbNlA1cyZ1n3yKsFpJuuACUn57NdasrH0a7+fyn7n2s2sZ6BrIi6e/SKzl0CtMrMSaQqFQKPZI6UMPUzN7Nr1fmEXC+PFttpNSUvtBPg0/lpJ0Th7awGRdoC0to3JbPQjI6pdE/7Hp5B2WTpyj/WSD9vDXhyj6RXeVbltbTTikYYuzkDNct7j1HpKM1Wbe7+N0JsGiInbccSf+1atbbBc2W5Nwa3ps/ry9bfEtt4mYmKiyNvpWrKByxkzqFy7EFB+P69LJJE+diiV135cbK3AXMGX+FFx2F2+c+QYuu6v9TgchSqwpFAqFok08Cxey/fobSJ56BRn33LPHtu7Pt+D5civeXCeraoOUFdYBkJHroP/YDPqOTifB1XUJAqFghG1rqylcUUHhL5UEGsKYrSZ6D04mb1QaOSNSiE3Yf4HYGchQiPpFiwjX1KA1NKB5vS0fG7xo3l0eGxrQGhogEunYQSyWVkRdHKa4xse2xZ85Ph4RF7fb496KPykl3iVLqJwxE+8PP2B2OnFNvYLkyy7D7HTuw5XbSYW3gsvnX44/4mf2WbPpnXjo1lJTYk2hUCgUrRKuqKDg3AlYMjLIefedVqvEA3jrgmyfm0/c2iq2BDVWeCOk9k6g/9gM+o1Jx9FKckFXo0U0ivPdFK6ooGBFBfU1AYSArP5JTe5SR8rB5y6TUiKDwWairpmQa75tN/HX9jYZCHTs4EJgio1t28rXJAT158Iag/vjj/GvWoUlLY3kq6/GddGFmOLj2z9WO9QH67nq06vYUreFV854haEpBy7LNhSMULG2iNLvv6W+oo7j/nRTlx9TiTWFQqFQ7IbUNLZdOw3vsmXk/uc9bP36tdjvbwhRsKKC/KVlRDa7GRtrospsInR0Fv3GZuDK3P8v5M5CSknltnoKDOFWXdwAQGrvBPJGpZE7Mo2UnvFR5TI8kMhQCM3n65jQa7L07S4SI94GZIMXzettGtvaqxcp11yDc9J5bYr9vSUUCXHjlzeytHQpz578LMf2PLZTxm0NKSWeKj+lhW5KN7spW7eNynLQpB4TmWSv4ZKHT8Ecv39WwvaIhjprCoVCoYgyql9/nYZvvyXzwT83CbWgP0zhykryl5axdW01WkSSnWJjbLwZc2Y8I24YgSkm+r4uhBCkZSeSlp3IkefmUVvupXBFJYUrK/jx40J+/KgQR6qd3FF6SZDMPCemQ2w5rD0hrFbMVitmh6NTxpOahvT50LxezMnJCHPnxQxqUuOB7x5gSckSHjrmoU4XauFQhIotHkoL6igtcFNa4MZr1PyziAAZ1o2Mcm4nc0g2mSecQWx2dK2EqSxrCoVC8SvBv24dRRddTPxxx5HxxFNsXV1N/tIyilZXEQlpJLhs9BuTTt+cRCIfFWBOspF+3QhMcR3P6IwWGtwBilZVUrCiku0bqtHCkthEKzkj9ASFXoNcWKzRlaDwa+bJZU/y8uqXufmwm5k2Yv9XmPRU+5tEWWlBHZXbPGgRXe84nBqZMflkBhaSad1AysA8TGOnwsAzwXxg73XlBlUoFApFE5rPx+bzL6ZcplN/7u8oWl9HOBAh1hFDv9Hp9B+bTmaek0iNn/J/rURYTPoyUgfJygJ7IugLs2VNFYUrKihaXUXIH8FqM5M9NIW8Uan0GZaC7SAUpIcKb617i0d+fISLBlzE/ePu32u3dSSkUbHN00KcNdTq8XoWq4n0HAeZWZAR/p7M4leJ822EhEwYPQUOmwKuPl1xWh1CiTWFQnHIoUU0KrbWY0+wkJBsj8qaW9GGFtHYvqGGVbM+ZYfHQdgahy3eQt/DdIGWNcDV5BqMeIKUz1iJ9IVJu34k1vToXklgX4iENLZvrNEzS1dW4q0LYjILeg50kTcyldyRacQnHfwC9WDhiy1fcNvC2zih9wk8ecKTmE3tWzvrawK6KDPizSq2edDCupZJTLGTmeckM89BZp94Uhq+wbzyNcj/EoSAfqfCmCuh/2lg7n7XvhJrCoXioEdqkog7QLjaT7Dcy6YF26iv8FEQ1PBJiE+y4UiNJTHFTmKKHUdKLI4UO4mpdhKSbJh+pWJO0yQl+bVsWlrO5uXl+OtDmMM+eiV5GX7lSfQa7NpN6Gr+MBWzVhGu8JF67XBs2Z0T5xTNSE1SVlTXlKDgLvcBekkSPUEhNaoSKg41lpct59rPrmVwymBeOO2FVoveRsIaldvqm1nN3NTX6FYzs8VEep9EMvKc9MhzkpHn0NeYrS6E5a/Dz7OhoRwcPXUL2mGXQ1J0lQFRYk2hUBwUaL4w4Wo/4WofkeoA4Wof4Wo/kWo/4doARHb+j9KkRJgECIGnVwIldivuaj+eKj/1tQFo9u/MZBIkJNuaRJz+aCcxVRd08U6bPtYhgpSSssI6Ni0tY/OychrcQSwxJrIHJJD44T/JTGig79uzEa1k7smwRuWrawgUuEmdOgT7wORuOIPuRUpJTYm3aQWF8i3G0leZcXqCwsg00vskHlL3THdSUFvAlP9NIdmezBtnvkGSPQnQYw0bXZllBW7Kt3qIhDQAElw2Mvs6ycx1kpnnJLV3AmaL8aMjHIQN/4Vlr0LBQhAmGHCGbkXrdwp0wGLXHSixplAoogIZ0YjUBgxBZoiwZn/SF27R3hRnwZxsx5Jsx5IcixZn4YeF2ykt83H0FYPIG5SM++MCfL9UYkmxkzShH/YBLiJhjfoaP3WVunirq/K1eO51B1sexyJIdNlxpNpJTIk1HncKuzhHdFWJb43G0hWblpaRv7QcT7Ufk0XQZ2gK/Q/PoM/QZEpumIZv5Upy57zf6tqMUpNUv70e36pKXBcNIH50RjecSfThqfZTuFLPLN2xsRapSeKTbOSOTCVvZBpZA6Nz6auDgXJvOZfPv5xQKMwzI2chy2x6luZmN55qP6B/PtOzdatZozhrtfBy1WZdoK14C7yV4MyG0VfAYZeBY9+WuTqQKLGmUCgOCFJKNG+4hQiLGJaycLWfiDsAWrMOZoHFZW8myPS/xtemZot9e6r9fPj0CjzVfs6YNoyc4TuXr/FvrKH2w82EK33EDkvB+Zu+WPYQaxQORfBUNYo3Px5DzDU+93lCLdqbrSbdEtdMwCWm2HEYljl7grXbxFx1cQObjAXT3eU+TCZBr8HJ9D88ndyRadhi9WtY9eKLlE9/nB4P/Y2kCy7YbRwpJbUfbqbh+xKcZ+WSeFyvA30qBwX+hhBbfqmkYGUlW9dUEQ7qS1/1GZbStPTVrovUK3bHWxekcGMpry14F3ulix6+XDTjYxefZNPjzPJ0YZbWOxGztQ0xHA7Auo90kVa0GIQZBp0Fo6+EvidGrRWtNZRYUygUnYYMa4RrdreKNb6WgZbL5JgSrLuJMP15LGZHTIdcSbVlXuY99TNBX5izfzeSrP5Jrc7Ls3g7ngXbAEg8OZvEY3siLHtv8QgFIk1WOE+Vn7pKX5Owq6vyEWhoaQG02Mw4Gl2rrVjmbHGWThVzteVe8peWk7+sjKodDXrV/gEu+o9Np+9h6dgTWmYz+n5ZTdHkySSefDI9n3qy1bnUfbWVuk+3kHBcT5LOyuu0uR7KhIIRtq+rpmBFBUWrqvA3hDBbTPQekkzuyFRyR6QSmxgdS191J1pEo2pHQ4tYs7pK3WoWEWESe1jpP6gXGYZAS0y2tz9oxUZY/ppuRfNVQ1IfGDMVRl0GiZldfEZdgxJrCoWiw0gp0epDugCr8ROuainIInUt48GwmLAk27Akx7YiyOyYYvbvl23FVg8fPbsCgHNuHkVaduIe24dr/NR+VIB/bRWWtFiSJvTF3q9zF34O+sK6cDNEXJOrtcqPp9JH0N9SsMbYzU3xcU0xc40u1xQ7MbHtW2I81X7yl+oLplds1WOoevR10m9sBn1Hp+nB1K2gNTRQOOl8tECAvHlzW12rseHHUmrmbCLusHRcFw5QsVj7gBbRKNns1uPcVlTiqfYjBKTnOIhNjMFsMWGxmjBbTZgt+qPFasJsEZgt5pavrcZriwmzteV+k0VgabbfZBFR56L31QebCs6WFbgpK6ojHNRN6nHOGDJyHSwX37Io+D9uPu1azh14TscGDvlg7Ye6FW3rd2CywqCz9Vi03OPB1LmuaJ+nji2rfqa6eAdHX3hpp47dGkqsKRS7oPn91C9YgK1/f2z9+3f3dA44MhQhXKPHjkWqfLtZyGRIa9He5Ihp1U1pSY7FlGDtsi/34k21/PefK4mJszDhD4eRlNHx8hG+9dXUfrSZSJWf2BGpJJ2dd0DqhEkpCXjDrcbKNVrpGr+4GrHFW3Zmrxru1UQj8aFkcy35S8sp2ewGIL1PIv2M9Tg7YoEovu8+3HM+IPu1V4k/4ojd9vvWVFE1ey32AS5SrhiCULFX+42Uksrt+tJXOzbUEApEiIQ0ImGNcEhr8ZxO+NptFH9mQ+xZrOad2yzCEImN24QhAs0t91vMu7dvY78uHM2YLQKTxYS73NsUZ1Za4MZdoWfSmkyC1N4JeqxZM6vZk8ue5JU1r/CH0X/gmuHXtH+CZWt1K9rKt8FfC8l5ukAbeSkkpO3/BTTQtAhlm/MpXLGMohXLKNm8EaQkNtHBtf98GautAxa//UCJNYXCQEYiuD/6iIqnnyFcUgJA3NixJF1yCYmnndpp69pFA5o3RKjc22QZa+621DwtA+xFjKnJNWlJtmNx2TCn7HwuuqG6e9EvlXwyazWOFDvn/H5Ux1wjuyBDGp5F26hbuB1hEjhOySbhmKxuFSRSSvz1oRaWueZxc54qP5FwSzGX0jO+SaAl7UW9s7pPPmHHLbeSct11pN96y277A4VuKl5ajbVHPGnXDt9vK6hi75BSommySbxFQi3FXCSkEQ5ru+3XGkVfeJfXuwjBSLgD44Y0OuurPzbR2hRnlpnnJK1PItZd7qk3173Joz8+ysUDL+a+I+9r2yoY9MKaD3SRtu0HMMfA4HN0kZYzXq+T1gk01NZQtHI5hSuWsWXVz/jrPSAEPfoOIGfUaHJGjiGzX39MByD2TYk1hQKo/+ZbyqdPJ7B+PfZhw0i96XcE8/OpeeddQtu2YU5OJun880m6+CJieh18wdURT5BAkZtAgZtgYR2hsoadv9oFmJ02zK5mgfwpzQL547svQL41Nv5UypevrCOlVwLn3Dyy1bifiNtNYNMmYkeMaLUERXPCVT7dNbq+GktGHK4JfbHl7R73Fg1ITeL1BHX3arWf5Kx4UrIS9nqcUHExBRPPIyYnh5w3ZyOsLePYQqUNlM9YhTnRStr1IzHHq6r9v1a0SDNxF5JEwpFm4k4SCUUMUSgJhyPNRKEkHIqQ4NKLzzpS7Xv8P/JZ0WfcsegOTux9Ik+c8ETrRW9Lf4Flr8GqdyHghpT+hhVtMsSn7Pe5RsJhSjaup3DlMopWLKe8aDMAcc4kckeNIWfkaPqMOIzYxANfW1CJNcWvGv/atZRPf5yG777D2qsXabfeguPMMxFGfIPUNBq+/Y6ad96mfsFXICXx44/FdcklJBx/fKcuUNyZhN0BggVuAoX6X9hwPQiriZg+Dmy5TmJ6JegWsiTbPgXadwerF21n0dsbyeqXxNk3jmgRzyXDYeq/+Qb33HnUL1iADAYxu1w4zz0H56RJ2AcO3OPYvrVV1H64mUhtgLjD0nGelYs5CgPAvXVu8n/8nsIVy+g9dASHnX520/3aEWQkwtapV+Jfu5bcuR8Qk53dYn/YWEYKIP2GkVhcXeveUSiWli7lus+vY0jKEF447QXslmb3XKAe1szRY9F2LAOzDYZOhNFToc/R+21Fq6ssp2iFbj3bunoFQZ8PYTLRc+AQckaOJmfUGNL75O7VZ6wriAqxJoRIAl4EhqH/3r8aOB24Fqgwmt0rpZzfSt8zgKcBM/CilPLR9o6nxJoitGMHFc88g/vDjzA7HKTeeANJkyfv0dUZKi2l9t33qH3vPcIVFVh69MB10YU4zz8fa3r6AZx9S6SURKr9hjCrI1DoJmLUIBI2M7YcB7Y8JzG5TmJ6JhyUcUdSSpZ9soUf5hWQMyKV068ZisVwofg3bMD9wVzcH39MpLISs8uF4ze/Ie6wUdR9/jn1X3yJDIWwDx2K8/xJOM8+u9VAegAtGMGzcBueRdsRFhOO0/qQMC4LYe5ey6K/vp5NP33Hhu8Ws3X1SqSmEetw4qtz02vIMM644Rac6R3LcqucMYOKp56mx6OPkDRxYot9kYYQFTNWEvGESL9+BFZVlV/RxeTX5HPFJ1eQGpvK62e83lT0luIVukD75T8Q9EDaIN2KNuJiiNv3YszhYJDt69dQtGIZRSuXU7V9KwCJKWnkjBpN7qgxZA8biS0uuu79aBFrrwGLpZQvCiFigDjgFqBeSjl9D/3MwEbgVGA78BMwWUq5dk/HU2Lt10vE7aZy1ixq3pgNQPLUK0i59lrMjo6btWUohOerr6h9+x0avvsOLBYSTz4Z1yUXEzduXJe7DKWUhCt8TVazYKGbiFHI1RRnISbHiS3PiS3XibVH/EGfvSel5Ls5m1nx+VYGHJnBSVcMRtbWUPfxx9TOnUdg3TqwWkk84XicEyeSMH58C9dnuKaGuo//S+377xNYvx4RE0PiqaeSdP4k/f1q5RdzqNJH7YebCWyswZoZT9LEvthyWhd4XUXA20D+T0vYuOQbilb+jBYJ48zIZOC4Yxl49HGk9cllzcIv+Oq1F5CaxvFTrmbEKWfu8f7zrVhB0WWX4zjjDLKmP9airRaMUPHCL4RKGkj77TBsuQf2fBW/Psoayrhs/mVEZITZZ82mpyURVv9HF2klK8Fih6GTdJHW+4h9tqLVlBZTtGIZhSuWsW3NL4SDAcwWC72GDCdnpC7Qknv2jqpwj13pdrEmhHACK4A82ewgQogHaV+sHQU8KKU83Xh9D4CU8pE9HVOJtV8fWjBIzZtvUTljBlpdHc4JE0j7/c1Ys/avcnWwqIiad97FPWcOEbebmJwcki65mKSJEzEndU7ck9QkoTIvwYJaAkW65Uyr1ytEmhKt2HKdTX+W9LiDXpw1R9MkC99cz7pvSxh2XA9GubZSN28u9YsXQySCfdgwnBMn4jj7LCyu9ktw+Neupfb9Obg//hjN7caS1YOkiefhnHTebrGIUkp8q6twf7yZiDtI3JgMnGfmYE7oOtdo0Odl87If2fD9NxStWEokHCYxNY2BR41n4FHjycjrt9sXSl1lOZ/NfJYtq36mz4jDOO263+NI3T0LLlJfT+F5kyASIXfeXMyJO8ucyIhG5WtrCWyqIeXyIcQO3f/4H4ViT3iCHqZ+MpXi+mJeHf1HBm38Cn55H0INkDFMF2jDL4TYvf8/GvL72bpmFUVG7FltmZ4wlpTZg5yRY8gdNYbeQ4ZjtR88Lv5oEGujgFnAWmAksAz4A3AncCVQBywFbpdS1uzS9wLgDCnlNcbrKcCRUsqbWjnONGAaQHZ29pgtW7Z0yfkoogupadT9dz4VTz1FaMcO4o89lvQ7bsc+aFCnHkcLBPB88gk1b7+D7+efETYbjjPPxHXJxdhHjtyrX2wyIgmV1BNojDkrqmtaasnstDVZzWJyHVhSY6P61+D+EAlpfPbyGgp+rmBQwhZ6Lp6JdLuxpKfrcWgTJ2Lr12+fxtYCAeoXLKD2/Tk0fPstSEncuHEknT+JxFNPxdTsn7gWjOBZsBXP4h0Iqxnn6X2IP7JHp4nikN9Pwc9L2fD91xQuX0o4FCQhOYUB445l4FHj6dF/YLvvsZSSVV/8j0VvvIwwmThx6rUMPeGUFv2K77oL90cf02f2G8SNHr2zryapeW8j3p/LcU3qT/wRB2fR0IMNKSW1ZSXEJjiwJ+x9ksjBTDAS5IbPrmV5+Qqe98dyVMk6sMbBsPNhzFXQc/ReWdGklFRt39pkPduxfg2RcBiLzUb20BHkGMkBrszoX1aqLaJBrI0FlgDHSCl/EEI8jS7QngMq0WPY/gb0kFJevUvfDou15ijL2q+DhiVLKP/HY/jXrsU2ZDAZd9xB/NFHd/lx/evXU/P229R9+BGa14tt8GBcl1yC8zdnY4rfPQ5ChjWCO3aKs+CWuqZK/5YUOzGNlrM8568m2Nu7pZj5zy2nzBNH//z/kF3xHYmnnIJz4kTijz6qUxM7QiUluOfOpXbOB4S2bcOUmIjjrLNIOn8S9uHDmwRPqNxL7bx8ApvdWHsmkDShL7bsfcsKCwUDFK1YxobvFrN5+Y+EAwHinEm6QDt6PD0HDN6ngObaslI+nfEU29euJm/04Zw67WYSXMm4P/qY4jvvJPWmm0i76Xct+8wvoP7rHThO74PjxOw2RlZ0JlXbt7Hg1Zls/UUv6GyLjycpowfO9EycGZkkZWTiTM8kKaMHiSmpmKI0kWmvkRJt2w/c/c29/C9Uwd/LKzknsd9OK5q945+ngLeBLb+s0AXayuXUV1UCkNIrm5xRY8gdOYaeg4ZgOURKLkWDWMsElkgpc4zX44G7pZRnN2uTA3wspRy2S1/lBlXshn/DRsofn07D14uxZPUg/dZbcZy9dxlznUGkvoG6jz+i5t9vE9iwAVN8PM4J5+I8/2KEPYNgY8zZVk9ToVlLepxhOdMzNs2Ori/UGi1oXi+eL76gYu58vvcfTl1iH0YEvmH4b4aReMYZmLvY+iA1De/Spbjfn0Pdp58i/X5s/fvhnHQ+znPPwZKSortGV1VS+98CNE+Q+MMzcZye06HSFuFQiC2rlrPhu8XkL/2BkN9HbKKDAeOOYcC48fQaMrRT6jVJTePnTz5i8b9fx2K1ctzEi7D+9RHsAwbQ5/XXEJZma6p+vR33/EISjs7CeU7eIWuljRaCPi/fv/82y+fPw2q3c/i5F2A2m6ktK8Vd3vhXjhbZuWSZyWzGkZq+U8Rl9CCpmaiLie14bb1uw1cDK9+B5a8xPVzMa04Ht8T25bfH/hmyDuvQEFLTKC8qaKp7VrxxHVLTiImNo8/wUU11z1oLATgU6HaxZkxiMXCNlHKDEasWDzwhpSwx9t+KbjG7ZJd+FvQEg5OBHegJBpdKKdfs6XhKrB2ahEpLqXj6Gdxz52JKTCT1+utxXXYpJlv3Cp6IP0z9Vz/jWfgLkRqBydkHYbYCEktmPPa+SbpbM8fRpfFQ0UiTQJo7D88nn+ALW1g55la89lROnpTFgFOHdMu8IvX11M2fj/v9OfhWrtSTSE48AeekSSSMH4+MQN0XW6n/dgcmuwXHGTnEj83czTUaCYfZ+ssKNny/mPyflhDwNmBPSKT/EUcx4KjxZA8d0WVWk+riHXzy/BOUbNpAZr2fs6c/S9LgwU37G5aXUfPuRmJHpJJ8yaBDKtYx2pBSsv6bhSx68xUaamsYfuKpHDt5KnGO3ZM4NC1CfVUVtWUlTSKutqwUt/HcX+9p0T420aFb5ZqscY1CrgcJruTuKzkRDsDW72HFv2HtXAj7eaP3YP5haWBy/wu456g/tfvjwFvn1q1nPy+laNXPeN21AKTn9m2qe9aj/yDMlvaXZDvYiRaxNgq9dEcMUABcBTwDjEJ3gxYB10kpS4QQWeglOs4y+p4FPIVeuuNlKeXD7R1PibVDi4jHQ9WsF6h+/XXQNFyXX07qddM6LcB/b9F8Yb0ArVFKI7TDAxpgAmtGLNJfjHfppwTWf4853obz/Em4Lr54t3pXhzLBLVtwz/sQ97x5hHbswBQfj+nUCXwfORZfUHDWDSPoPWjf0/M7k0B+PrVzPsA9bx6RqirMaakkTZiAc9IkTHEZ1MzLJ1hYh7V3Iq4JfbH0iGPrmlW6Be2n7/HXe7DFxdPv8KMYeNSxZA8fdcC+XMqeeZal773Fxl5p2OITOOXa3zHgyGPwbaim6rW12PKcpF459KCps3cwUl5UwIJXZrJj/Roy+/bnpKuvp0e/Pdf82xP+hnrcZaU7rXFlpdSWleAuL6WusgKp7Vzhwmyx4Ehv7lZttMxl4MzI7NwlkqSEivWweQFs/gq2fAshL9gcMOIiPskayB9XPcfJ2Scz/fjprRa91bQIpfkbKVyxnKKVyyjdvAmkxJ7oIGfEYXrds5GjiU/q3PV8DwaiQqwdaJRYOzSQwSA1b79D5fPPE6mtxXHuOaT9/g/E9Op5QOcRqQ8SLKprijkLlRqrA5gFMb0TmzI1Y/o4MNn0f1BS0/AuWULNv9/Gs2ABRCLEH3MMrsmXkHDCCS1cVYcKEY+Huv/9D/fcefiWLwchiD/qKJznTSQ49Gg+nrmOSFjjnJtGkZF74CuEt4cMhahfvJja9+dQv3AhRCLEHnYYzkmTMPc8HPdn28CvUeRbw8/lXyBsJvqOHcfAo46lz4jRWKwHdhUA77JlbJlyBc5zzsF28438759PUlawidGHn01/9wis6XH6MlL2Q+9eiwb89fV8++5sVn42H1tCAuMnT2X4iad2qaUrEg7jqaygtrwU926WuRKCPl+L9vFJrhbWuMY4OWdGJvFJrvbd4p4yKFgIBV/pjx4965KUfpB3IvQ9CfKO56fqtVz3+XUMTx3+/+ydd1xT5/7H3yeEHYYsFZClgIoK7ol7b604Wm1rW1s7bXu7x723ve399ba93Zw08KYAACAASURBVLa1tWrnrXu07i3ujVtQ9gZJIIFA1vP74yBKRQVNHG3erxevkHOedSDJ+eT5LuYMmlMr6a1OXUJ60mHSL5Z0KtchSQqatIgkPK4TYXEdaBzR4paUdLqTsYs1O3cdQgi069ZR+NHHGLOycOvejYAXXsA1JuaWzG8uq7oUqZlWiqnwsuoAIR7VkZpeOId41KtuprGgAM2SJWgWLcZUUICycWO8ExLwTpiAY+PGtr4cmyLMZsp376Z0+Qq0mzcjqqpwiojAa+xYvEaPwrFJE/LTSvn98ySUjgpGzYq7ofJJtxpTURGalas4v3I5WRWl5HursDi60ta3Dy082oOThNfwcDy6BN0W86K5rIy0sePAwYHw5ctxULljNpk48usKGiV5YsKAcqwfzXt2u+VruxpCCC5UXiBHl0OuLpccXQ4WYaG5d3OivKMI8ghCId35O4DCYuHEtk0k/rKASp2O2MHD6DFxKq4qj+t3tuW6hECvLbtsNy6/1u/akmIuLwSqdHLGq3oH7mLwg7efD16mXLxKDqPM3AEFJ+TGrj4Q0adanPUD72Y146SoU3hg7QP4u/nzw7AfUDm4kZt8pqYgelFGGiALx7DYjoTFdSC0bdxtKel0J2MXa3buKsr376fwgw+pPH4c56goAl58EfdePW3qGG2qqQ4g/5gvXKoO4BTqeSmVRpDqpsxJwmRCt3076v/9SvnOneDggEf/fnhPnox79+63vdxJQ6hKSUGzYgVlq37DVFSEg5cXniNG4DV2TK0Iy6zTJaz5+jhunk6MmRWHp5/rbV75tRFCkJdyhrN7dpK8dye6kgs4KJUEuqjwT0nDv0iNa3RHXGKnIqrccAr1xHtMc5xuoQAVQpD7t79Rtn4DYb/8jGtsLCCXICv6Mgmzwcge3SoyM08S02cAfR+YgYu77dcnhEBTpSFXl0u2LrtGkF0UZ7m6XCrNlVft76p0pYV3C6IaRRHZKJJI70giG0XSyOXOMYnln0tm8/yvyT+XTGB0awY8NJOAsIjbvax6YTIaKSsqkMXb5TtzOeloioowmSyXtRaoXCS8fbzxCo7AO7QVXk2a1uzKuXp4IkkS+eX53LfmPlx08LL/DNRnzpF5IgmDXo/CwYHA6FY1ec/8Q8PtAS7XwC7W7NwVVKWkUPjfj9Bt24aySRP8Z83Ca/Qom9TmFGYLFUeLqDqnkcWZpgoAyVVZU7pJrg6gslkZIkNmJppFi9AsXYZZrcYxNIRGEyfhNX5cvZK/3g4uVgooXbGCypMnQalE1bs3XmPHoOrb94pSXuePFLLhu5M0auzGqGficPe6MyNfhRAUnE/h7N6dnN2TiLa4CAelkrC4TkT3iKd5h844ubrJ0awbN6JZuoyK/QdQhvbAJXYSksIZ925N8BoacUtMjprlK8h79VX8n30Wv5mPAWCpMFI45xhmTRX+j7bDobEze5f+yr4Vi3Fv5MOQR58mLK7jTc9dZigjR5tzhRC7KM4qTBW12ns6eRKkCiJIFUSgKpBAVSDBqmACVYEEqWR3hnOac6SoU0jRpMiP6hTUVZdSbvq5+tUIt4s/zb2a164vaWMqykrZ+esPHN+yAXcvb3pPfYhWvfreneKjNFv2OUvdCqnboaIYIaCiUQwany6Uujan1OKJpvhCjYm1XF1SawjhpKBSJaF20qPSKfDUya97Dz9/OTAgriMhMbE4u90Fkax3CHaxZueOxlhQSPEXn6NZugyFmxu+jz2Kz7RptZKWWgshBJVn1ZSuTsVUpEehulQdwCncC8fGt746gMVgQLt+A+pff0V/6JBcKmnoEBpNnoJr+7jbfjMQBgO6HTvQrFiBbvsOMBpxbt0K77Fj8RwxAqVv3ZnwT+/OZeuPZ2gc7smIJ2NxqUfqi1uJEILC9FSS9yRydu9OSgvyUTgoCYttT3T3eJp36nrN2oGGzEw0y5dTumodDr7dcQzvg6Qwoerli9fwhiVJbgiGjAzSxo3HJSaGkAXzkRwcEEYzRXNPYMjW4je9DS4tLgXe5J9LZu2XH1OSk0W7AUPpM+2ha6aCKDeW1zJT5uhyZHFWnkuONgetsXakoruje40Qu1yUXRRkHk4NNw1eNJcmq5NrxFuKJoXzmvNUmeUvVgpJQYhHSK0duMhGkQSrgut0bL9RLBYzxzauY9fCHzFU6mk/bDTd75lyd4mQKi2k77wk0IqT5eOqxhDRt9q02Rezuz955XlklGWQXpZORllGzU9hWT5u5Qo8KpR4VCjxN6jwq3LHvcIB/4Bg2nXpS1hsR3yCgm/7Z9bdil2s2bkjMet0XPjuO0rmL0CYzfjcOwXfmTNttqtkLChH83sqVSkalH6ueA0Px6WVzx31wVJ5NhnNwoWUrlyJpbwc5+hoGk2ehOeo0Tiobl3RYSEElSdOUrpiBWWrV2PWaHDw88NrlFxVwCU66pr9j27KZNeSczRr7cOwx9ri6HznOA4XZ6Zzdk8iZ/ckos7LRVIoCG0bR3T3eFp07t7gTPPCbKZ8z140yzdj1oXj0CgMS0U2brEONLpnsFUjloXRSPq992HIyCBi5QocmzZFmAUXfjpF5ZkSfKa0xK3dlTmoTAYDuxb9xMHfl+Ph50fstElUBbnVFmTVAk1TpanV11XpSqB7YC0xFuRxSZB5OnnesveQ2WImU5t5xS5cljYLgXz/cnFwobl38ytEnJ+rX4Pnyzlzis3zv6YoPZWQNu3oP30mvsF3QUS32QS5R2Rhdn4rZO8HiwmUroiQ7hSHdSPdP4IMBWRqs2qEWZY2C6PFWDOMu6M7oZ6hhHqGEuYZRohnSM2jp5Pd38za2MWanTsKYTSiXrSI4tlfYi4pwXP4cPyfexanZs2u3/kGMJcbKduYQfn+PCQnJZ4DQ1B1a3pHpzKwlJdTuno16l9/perUaRRubniOGkWjKZOtXkbrcowFhZT9tgrNihUYzp1HcnJCNaA/3mPH4t6z53UjWIUQ7P8tjYNr0mneIYBB01vj4Hj7/84XcrI4uzuR5L07uZCdiSQpaBbTVhZoXbrXmQvrRjCp1ZT8uJ2qLDeQnDCmb8MpuALv8WNw797tpk36hR9/woU5cwj65BM8hw5BCIF6aQoVBwvwHtMcVfdADGZDjX/Y5X5jubpcyjNyaXfACa8KR06FlnE4WoPCybFGiNUSZNXPfVzurC80dVFhrCC1NJUUdYq8G1ct5EoqL5nufFx8aptSvSNp7t0cN8crd8h06hISf57PqcStePj60/f+h4nsalu/2ZumJE1OqZG6FdJ2UGooI1PpRLp/OBm+IWQ4u5FhriBDm1nLVO2kcCLEM4QQjxBCvWRRdlGg+br43tnX/CfDJmJNkqTmgJsQ4vjNLM5W2MXanYcQAu2GjRR99BGGjAzcOncm4KUXcW3b1jbzmSzo9uRRtjkDYTDj3rUpngND65WJ/k5BCEHlsWOof11I2Zo1iKoqXOPi8J48Cc+hQ61iKrZUVqLdtJnSFSso370bLBZc4+Lk4unDhuLgVT8hIyyCxIXJHN+eQ+ueTelzX0sUtzEJqzo/VxZoexIpykwHSSK4ZQzR3eOJ7NrDpnmczOVG1AuPUpmsx1JVRtWxRWDJxnvcWLzGjbuhLybl+/aT+eCDeI4fi+WVx8nR5SBt0xCU5M7+FimsCNpGjjaHQn1hrX5KSUlTVdOanbCmzgE47sxCvfs4HgEBDHvyeZq1bHOVWe9uLugv1NqBS1GncL70PHpTdXQ3EsEewTUBDS08IpCO5HL293WYjUY6jRpP17ET78xi4Ho1+nObyDy3joy8A2RWqUl3VJLh4k6GkxNqcWmHTCEpCHQPrC3GPEIJ9QqliVsTq5qN7dw4VhdrkiS9BrRATgPqLISYdnNLtD52sXZnUXH4MIXvf4D+6FGcWjQn4IUXUPXpY5NvbUIIKk+XULomDVOxHueoRniPCMex8a0zI9oCs0ZD6cqVqH9diCEtDQcvL7zGjaPR5Ek4hYU1aCwhBPrDh2Uz59p1WHQ6lIFN8RozBq/Ro3EOD2/Y2swWNi84TcqBAtoPCqH7+Oa35Rt5aWE+Z/fIQQKFaecBCIxqRXSPeKK69kTlU7d/na0wZGlRL0/BmFuOMOZRseNrLNo83Lp2vVRQ3rXu6FiD2cD27O2kqFMoyk9j2D/Wo1cKXpwuoXcUjC7py+MFE1nnvYvFzbcS5BlEoHsgQR7Vu2LugQR7BOPv6l/nzTjr1HHWf/UJpUWFdBo5jp4Tp/5paixeC7PFTI4uR96F01zyiTOkFdDlpDfeOidyA6oo7OZFSGi0LOS8L5lSb/Xr2mgxkqvLJUN9nvSsnWTkHyZDm0mGRU/+H3a6A5x9CPVuXmOuDPWUBVmwKhgnhz////Zu56bFmiRJzwCzhRDm6ucLhRCTqn8/JoRoZ80FWwO7WLszqEpNpfCjj9Bt2ozS3x//Wc/gNXaszRLCGvLKKV2dStU5DUp/V7xGRuAafWdkybcWQggq9u1D/etCtJs2gcmEe4/ueE+ajEf/fkjXSMxqyM6hdOUKSleuwpiZieTmhufgwXiNHYtbl843lDrEZDCz7tsTZBy/QLexEXQcGnYTV9dwyoqLaoIE8s/JjtNNWkQR3T2eqG69bnsdQWERlO/Po3RdBsJgwkFVRPmOeRjTz6NQqS4VlG/XDkmSSC9NZ0nyElaeX4mmSoMk4NVVStqerWLdq31wiYmhbWE4EVs9kaLcCJjWFifHG7sRGyr17PhpHkkb1+ITGMzQJ5+7qcz7dyNlxUVs//E7kvfuxNXXB69hHcltXEWK5hwpmhSK9cU1bb2dva/whYv0jqzTlNoQLMJCYUVhjTN/elk6GaUZZKhTyKkowMSllBqeZgthkjOhqiBCA2IJDe5WI9DcHe/uL6R/dawh1u4DHgQ+F0KskiTpYSABUABJQogXrbheq2AXa7cXU1ERRbNno1m8BIWLC74zHsHn/vtR2CiCyqwzULYhg/ID+ShclXgODMW9axMkh9vvL2VLjIWFlC5bhnrRIky5eSj9/fFOmIB3QgKOTZsCcrF57fr1lK5YQcWBAyBJuHXtitfYMXgOGoTC/cY/4Kv0JtZ8eYzccxr6TImmTe9bV1ki7egh9i5bSO7ZU4BcSzC6ezzR3ePxCrjzEg2bdQZK16ZTcagABy9nXFtZKN+7Eu369Qi9HkNIY3a3d+Hn0GzKVY70C+lHQlQCkYkZFP3jLQJefAHfhx+mMkVN8YKTODXzwP/hNvVKynw90o8dYf3Xn1KuLqHLmAl0u2fKLa/GcKsxGY0c+n05e5cvBIugy7gEOo+654rdRXWl+oqAhhRNSo0pFSBIFVQj3C6aVEM9Q1EqLn0pvZiDrq5Iy8yyzFr551xQEGoyE1pZQajJRKijN6FNOxHWfAjekUPA9faU2bNjW6xiBpUkyQV4AegC/B25uLqjEKLUWgu1JnaxdnuwlJdzYf4CLsybhzAYaDRpEn5PPH7V9A43izBZ0O3KpWxLJsJoQdW9KZ4DQlC4/blvNH9EmM3otu9AvfBXynckgiSh6tcPhbsb2o2bEHo9TqGheI0bi9fo0TgGBt70nHqtgd8+T+JCto6BD7UmstOtEUilhQVs++Fbzh3Yi3fjprTpN4io7r1o1OTmr+lWUJVeimbleYx55YgIFzaE7Ob8lp/pckhLdA5YHBS4xPcgYOIUnJo1I23iJFzjYgn57juMeRUUzTmG0scZ/8diUbhab4e6qqKcrd9/y8ltm/ALCWPYk8/fNcleG0rakYNsWTAHTX4ekV160Gfaww0S+BZhqTGlXi7kMsoyMMsGKBwVjkR4RdDMoxkFFQVklGVQZiirGUMpKQlWBRKq9CDUUEWoOofQC+mEGk0EOKpQRPStLufUDxqFWfcPYOeOxFpiLQYwAmXAv6oPvymEyLfKKq2MXazdWoTJhGbJUoq++AJzcTEeQ4YQ8NyzDfalqvd8QlB58gKaNWmYSypxaemD14hwHP3votxHNsKQnY1m4SI0S5ciTCY8hw3Da+wYXOOsl7NNW1LJqk+PoiupZOhjbQltY3tfMJPBwMHflrFvxWKQoNv4yXQcMfau2wEyWoxsy9hGxpbjxJ9vg5NQcjDsHIFDWtHB1ISy5bKZ2lxcDJKEg5cX4StXIDl4Uvh1EpJSQcATsTh42ibB8PlD+9n4zefotWV0Gz+ZLmMTbllRelujKchn2w/fcv7gPhoFBtP/wUcJi+1gtfGrzFWklabViLhkTTI52hyauDepduoPIdQCYcUZBGYdQpm5B0yVoHCEZl1kYRbRHwLjwO70/5fDGmbQBchCzQ3IEUK8JElSe+Bt4IAQ4m0rrtcq2MXarUEIgW7LFgr/+xGG1FRcO3ak8Ysv4BoXZ7M5DTk6NL+nYkgrRdnYDe8REbhE3ZkZ/28nwmwGIazuH6jOL2fVp0cxVJoZ8WQ7AlvY3iSTdvQQW+Z/jSY/j6iuPelz/8N4+gXYfF5rkqPLYWnyUpafW06xvpgm7k24N3gSg893QhzX4uDtjPeo5ri09gGTCV3iTsrWrsV73Fhc2nai8KskRJUJ/5mxNv9Sotdp2TLva87s2k7jiBYMfeI5/JqF2nROW2KsqmT/yqUcWLUEhcKBbvdMpuOIMTgob4HQL8u7lO8sdRuUV0fr+re8tHMW2hOc7/x6uXZsizXEWpIQIrb69yNCiPaXnRsjhFhptdVaCbtYsz36o0cp+OBD9IcO4RQeTsALf0PVv7/NoqXMWgOl62WfH4WrEs/Bobh3bmqzclB2rqQoU8tvnx8FYNQzcfg3s23h6stNno2aBtH/oZmEtWt//Y53CCaLiR3ZO1icvJhdObuQJIn4oHgSohLoFdSrJkqzKrUU9cpzmAoqcGnpg/eoCJS+cpSopdJE0ZxjmC7o8Z/RDicb/80vJ2XfbjbOnY2hopweE6fSadQ4FHfRjo8QgnMH9rDth7mUFRXSsmcfek+djodPwxPkNojyYji5HI4vgay98jF3f7laQEQ/+dHr1vl32rk7sIZYew9oDzgCa4UQH1h3idbHLtZshyE9ncKPP0G7fj0Ofn74P/UU3hPusVmEpzBa0O7KQbslC2G2oOoRiGf/EKv669i5PrkpalbPPoaTm5Ixs9rj3dh2uzvXNXkKAQfnwcH50O1xiJ0CNxDJaivyy/NZmrKUZSnLKKwoJMA1gPFR4xnfYjxNVU3r7CPMFnS7cynbmImwWPDs2wxVryAu/HCKqvQy/B6MuS07yBVlpWyaO5uUfbtpGhnN0Ceewycw+Javo6GU5GazZf4cMo4dwa9ZKP0fmkmz1rbJ6QjIJZ3OrIHji+XktMIM/q2g7T0QOQQat7mjXqN27jys5bPmCViEEDprLs5W2MWa9TFduEDxl1+hXrgQyckJ34cfwvfBB28qmvBaCCHQHy+mdG0aZnUVLq198RoejqNf3Xmp7NiO9OPFrPvmBJ6+LoyeFYeqke2ShF7u/F2nybNKC7/NghNL5d2K8iJoGgtD34PQHjZb1/UwW8zsyt3ForOLSMxJRAhBj6AeJEQl0Ce4T63IwGuOU1aFZnUa+qQiJEcFwmjBZ1I0bu1vn9lXCMHZ3TvYPO9rTFVV9JryAB2GjbqhVC+2xlCpZ+/SXzm0eiVKJyd6TppK3OARKG6yekSdmAxwfrMs0M6sAZMevJpB2wnQNgEax1h/Tjt/WuzlpuzcFBa9npLvv+fCt3OxVFbiPTEB/yeeQOlvu/xVhmyt7JeWXoZjEze8Rkbg0sLul3Y7SD6Qz+b5p/FrpmLk07G4qmyTXLNeJs+Ck7DofihJhX6vQ6/nZNG26Z9QlgOtx8Cgt29p9FxhRSHLUpaxLGUZeeV5+Lr4Mj5yPPdE3UOQ6sZNXZXnNJRtSMetfQCq7ndGpKtOXcLGb78g9dB+glrGMPTxZ/FuUvdO4a3moqDc/uN36NQlxPQdSPyUB6xfqcJigczdskA7uQIqNeDqAzHjoN1ECO5i30Gzc0PYxZqdG0KYzZQuX07RZ59jKixENXAAAc8/j3OE7cL5zWVVlK7PoOJwAQo3x2q/tCZIt7Fs0V+ZE9uz2f5rMkGR3gx/vB1ONjA915g8ly8ChXT1KM8jP8Hqv4GLF9zzHYTHXzpnqIDdn8OuT+SC1d2fhF7Pg4ttik1bhIU9uXtYdHYR27O3YxZmujXtxsToifRt1hdHxd0VodoQhBCc2rGFrQu+wWI20/u+6cQOGnZbd9mKMtPZMv9rsk+dICC8OQMeepzAKCvW0BUC8o/JAu34UtDmgqM7tBwh76A17wcOf97/uZ1bwx0h1iRJ8gbmAm0AATwEjAdGAQbgPDBdCKGpo286oAXMgKk+F2MXazdH5alT5L78MlUp53CNjSXgpRdx69jRZvMJoxntjhy027MQZoGqVxCe/ZqhcLH7pd0OhBAcWpvBvlWphLXzY8iMGJRWSL76R9KOHGTL/DloCi6aPB+5suKAoRxWvwBJv0B4b1moqa5iEizNgc1vw7FfwT0A+r8B7adaLQ1Csb6YFedWsCR5CTm6HHxcfBjTYgwTIicQ4hlilTnuFrQXiln/9adkHDtCSJtYhjw+65ZH6FaW69iz+BeOrP8dZzd34qc8QJv+g6wXBHHhvLxze3wxFCeDQgktBslmzuhh4GSvGGDHelhVrEmS1AMIA2ruokKIH+rR73sgUQgxV5IkJ+Q0IF2ALUIIkyRJ/6ke6+U6+qYDnYQQxX88dzXsYu3GqUxOJnPa/UiurjR+9VU8Bg+yWYSnEAL9sSJK16Zj1lThGiP7pV2MhLNz6xFCsGvpOZI2ZRHdtQn972+JwsqVIGqZPAOD6T/9sbqjPIvOwqIHoOgM9HkJ+rxcP+GVfQjWvwpZ+6BxWxj6f7V34hqARVjYl7ePxcmL2Zq5FZMw0aVJFxKiEugf0v8vXXNRCMHxzevZ9uN3SBL0vX8GbfrZ7vOiZl6LhZM7tpD4ywIqykqJHTiMnpOm4uphhZ1UbUF1JOciyDkkHwvtKe+gtR4Dbn+u8nV27hysJtYkSfoRaA4cRd7lAhBCiGeu08+ruk+EuMokkiSNAyYIIe6r41w6drF2SzCkp5M+bRoSEqE//4RTiO12CwxZWjS/nceQqcWxqTveoyJwjrCXUbmdWMwWtv58ljO782jbL5j4hEirmqD/aPLsfs+Uq+e7OrYIfnsWHF3hnm+hef+GTSYEnFwGG/8BpVnQciQM/hf41M+MX1JZwspzK1mSvIRMbSZezl6MaT6GCVETCPdqWKH7PzulhQWs/+oTsk4dJ7x9JwY/+jQqH9skSi5IPcfmeV+Rl3KWplEtGTB9Jo0jWtzcoJWlcPp3eQctbTsICzRpKwu0NveA150f/Wrn7seaYu000Ppqgusa/eKAb4BTQCxwCJglhCi/rM1vwEIhxE919E8D1Mjm0zlCiG+uMs+jwKMAISEhHTMyMhqyzL88xtxc0qdORegrCf3xB5xb3OQH4FUwlVZRti6diiOFKFSOeA0Jw61jY7tf2m3GbLSwYd5JUo8U0XlkOJ1HhFl1h6SWybNbL/pMe7juIuvGSlj3MhxaACHdYcI88LwJB3ujHvZ8AYkfg9kgp/ro/YLs+/YHhBAcLDjI4rOL2ZS5CaPFSIeADiREJzAodBDODrapGvBnQFgsHFm/msRfFuDgqKT/9Jm06tXXaq8hvbaMnb/+wLHN63Hz9KL3fdNpHd/vxn3ljJWQskEWaMnrwVwlB6a0TYA2EyDAij5vduzUA2uKtcXAM0KIvAYuoBOwF+gphNgnSdKnQJkQ4s3q868DnYDxdQlBSZKChBA5kiQFABuBp4UQO641p31nrWGYiopInzoVc4ma0O8X4NK6tdXnsBjM6HZko92ejRACj17BePQLRuFs90u73RgqTaz9+jjZZ9T0SogkdkAzq41dWljA1u+/5fzB65g8QfYRWvwA5B+Hns9C/zfBwUqvD20+bP4XHP0Z3Hyh/+vQ/n5wUKKp1LDyvLyLll6WjoeTB6ObjyYhKoHm3s2tM/9fBHVeDuu+/ITc5NO06NyNgY88eVMRmRaLmeOb17Pz1x+pqiin/dBR9Ei4F2e3G/AXs5ghPRGOLYbTq6CqTE7/0uYeWaQFdQQbm3Dt2Lka1hRrW4E4YD9QdfG4EGL0dfo1AfYKIcKqn8cDrwghRkiS9CDwGDBACFFx3UVK0j8BnRDiw2u1s4u1+mNSq8m8/wEMOTmEzJ2LWwfrZogXFkFFUhFl69IwlxpwbeeH19BwlD62y9Vlp/5Ulhv5/YskCjO09L+/JS27WScVQ4NMniCnQVj5lOyTNv4biBpilXVcQe4RWPcaInM3R5pEszi0DRuKkzBYDMT6x5IQlcDgsMG4Ku1+kzeKxWLm8OqV7Fz4I44urgx8+Amiu/dq8Di5yafZPO9rCtPO06x1W/pPfwy/kLCGDSIE5B6WqwmcWAq6AnDygFaj5ECB8D7W+0Jgx85NUF+xVp9X6z9vZAFCiHxJkrIkSYoWQpwFBgCnJEkaCrwE9LmaUJMkyR1QCCG01b8PRq5JascKmHU6smY8iiEjg2bfzLG6UKvKKEPzeyrGLC2OQSp8prTEOexK85Od20O5popVnx2ltFDP0EfbEBFnndx59TZ5ApiqYMObsH8OBHWChPngbTtfyVLfCH7vdh+L3Q2cryxElb+PcUpfEuL/Q3TEQJvN+1dCoXCg06jxhLfvzLovP+L3T94jZV88/R+aiZvn9d//5Ro1ib98z8ntm1D5+DJi1ktEd49vmEm1OKU61cZiOS+fgxNEDpZ30KKGyL6Qduzchdg6dUcccuoOJyAVmA4cAJyBC9XN9gohZkqSFAjMFUIMlyQpAlhefV4J/CKEePd689l31q6PRa8nc8YM9EeTCP78Mzz69bPa2CZNJaVr09EnFaHwcMJraBhu7QPsfml3EKVFFaz69Ch6rZHhT7QjOPrmk4f+0eQ50bst9gAAIABJREFUYPpMQtvFXb2DOgMWPyjvfHR7Aga+BUrrR1cKIUgqSmJx8mLWp6+nylxFG982JLQYy9DCDNx2fiZnn+/yGPR5EVztCZithcVs5sCqpexe/AsuKhWDZjxFi87drtr26Prf2bXoZ0wGA51GjqXr+Ek4udRTWJXlXkq1kZcESHK6l7YJ8k6aqz2Ayc6dizVqg+4UQvSSJEmL7ORfcwo5GtQ2mSdvArtYuzYWg4Hsx5+gfM8egj78AM/hw60zbpUZ7fYstDtyAPDoHYRHn2YonO+e4s9/BS7k6Fj16VEsZsHIp2NpHHZzb2GTwcCB35ayf/liJIWCbvdMvrbJE+TyPCtmyp8oY76A1tf0prghtAYtq1NXszh5McnqZNyUboyIGEFCVAKtfFtd1rAAtr4Dh3+UhVq/16DjdLt5zIoUZaSx9suPKUpPpXV8P/o9+BguKlXN+ayTx9gyfw7FWRmEtmtP/+mP1a8GqV4Np1bKZs70nYCAwPayQIsZD553RoUFO3auxx2RFPdWYxdrV0cYjWQ/9xy6TZtp+u67eN8z/ubHtAgqjhRSui4di9aAa6w/XsPCUHrb/dLuNPJTS/n9iySUTg6MfiYOn8CbS+yZeuQAW+d/I5s8u8fTd9rDePj6Xb2D2Qib35IrDjSNhYQF9U6nUV9OFp9kUfIi1qatRW/S08qnFROiJjAiYgTujte43rxjsP412QndvyUMeRda2E2j1sJsMrJ32SL2LV+Iu5c3g2fOwq9ZKNt/msfZ3Tvw9A+g7wMzaNGp27VNnoYKSF4nC7SUDWAxgm+LS5GcfraJZLdjx5bYxZqdGoTFQu7Lr1D22280fv11fKZNvekxq9JL0fyWijFHh2MzD7xHRuAcesdtttoBsk6VsObrY7h7OTN6Vhyefjfut3O5ydMnMJj+1zN5glxhYMl0OVltp4dhyL/B0TqCvtxYzpq0NSw+u5jTJadxVboyLHwYCVEJxPjG1N/fSQg4sxo2vAHqNNnPafC74B9llXXakXOlrZ39EReyM1E6OiEQdBkzgc5jJuDodJX0KGYTpG6rLpr+Oxh0oGpSXTR9AjSNs0dy2rmrsYs1O4Dst5P/z7fQLFyI/3PP4ffYozc1nqmkktK1aeiPF+Pg6YTXsHBcY/3tfml3KOcPF7Lhu5M0auLOqGdicfe6sZxhN2TyBEjZBMtmyLnORn0q32CtwOkLp1mcvJjVqaupMFUQ2SiSiVETGRExAg8njxsf2FQF+7+B7e/LJa+6zJArKNgz2FsFk9HI3qW/UlqYT89J0/Bu3OTKRkJA1v7qounLoaIYnL1kk3m7iXJlAWuVlrJj5zZjF2t2EEJQ+J/3KVmwAN9HHyXg+edueCxLlQnt1my0O7ORJAmPPsGoegejcLJ/aN6pnNqVy7afztA43IuRT7XD2e3Gik432OQJ8o7Itv+DxA8hIAYmfg9+kQ2eW2fQkaPLIVuXTbY2mxxdDseKjnHywkmcHZwZEjaEhKgEYv1jrVvuSFcE2/4tJ+l19oS+r0Lnh+2Fu21J4elLkZyaTFC6QNRQ2cwZOQiU9uTEdv58WDPP2tPAT0IItbUWZyvsYq02RZ9/QfHs2TSaOpXGr792zZuZyWhm/6o0zuzNQ5IkHBwVKB0VOCglmlgEIXoTjkKgdnek0N8N4aKU2ygVODgqcKh+VF72u4Oy+nld5y9/Xn1M4SDZvL7gX4UjGzPZvfQcITE+DH20LY43EOxxQyZPkBPRLn1E9gFrPw2GvQ9ObnU2NZqN5JXnka2ThdhFQXbxUVOlqdVe5agizDOMkc1HMjJiJF7ONk4JU3BS9mdL3Qa+kbIJN3KQ3fRmLTSZ1ZGcS6DgBEgKiOgnC7SWI8DF7lph58+NNfOsNQYOSJJ0GJgHrG9o6Sk7t54L8+ZTPHs2XuPH0/i1V68pgoqzdWyaf5ILOeVEtPfHxd0Rs9GCc1kVTYorcDVa0CkVnHFxohQw51VgNlowmSxYTBZMRgtmo+XmFy1Rt/i7TOD9UfzVCMbLn9clGJW1RSESSJKEdPFRIT8igcQfnl9sp7jUnj/0kyTg8vEutlP84bmNEUKwb2Uqh9Zl0KJjAAOnt8ZB2bDSPH80ecbf+2D9TJ4AqdtloValhbFfIWKncKHyAtmFybIg0+bUEmYFFQVYxKXXjlKhJNA9kGCPYFr7tiZIFUSwRzDBqmCCPYLxdPK8tYK+cQxMWyGXJtrwOvySAM0HyEEIAa2u39/OlZTlwtk1skDL3CMfC+4Cwz6AmLGgCri967Nj5w6kXmZQSf50HIycJ60TsAj4Tghx3rbLaxj2nTUZ9a+/kv/Pt/AYNpSgDz9Ecqh7V8ViERzdmMm+Vak4uzvSf1pLwtr6Ybqgp3RNGvqTF3Dwdpb90tr5XfMmKYTAYhKYL4o3kyzgzJeJuVrP6zpffcxkqmf7i8etKRhtzBUiDkBxpciTFNQp/mqOc0k8Xt7OYragzq+gdXwgfaZEo2igL2Hq4QNsXdAwk2e5sZzsskyy980m5+xKclS+ZAfFkmMoJUeXQ6W5slZ7f1f/GhFW61EVTIBbAA7X8EfKVlew6GA2CR2DaeZT926dzTAZ4MBc2P4eVOmg03To+xq426Z4+Z8CIeDCOcjYLQuzjN2gqa7f7N/yUtF0n/Dbu047dm4TVvdZkyQpFlmsDQW2At2AjUKIl25modbELtagdNUqcl9+BVWfPgR/9imSU93JRsuK9WxacIq8c6VEtPen733ROCskyrZno9uZg+Qg4dG3GR7xQUiOd4dfmhACi1nIgu8aAk8IgRBcerQIuPy5EPKP5dK4wiI/8od+F9vXHLdc6g9c9vyydpY/zC8E1NXuD+ukjn7CAvyhXdMWXrQfFNKgHajSwvxqk+c+2eT50ExC28omT6PFSL4uX/Ybq94du9xUqa6q7SHhrnQj2KPZFYIsWBVMoCoQF+WNRYIWlFWS8PUeMksqcHSQmNotlKf6tcBXdYt9mcovyP54B+eBkwr6vgydZ9gkse9dh9kEBcchYw9k7obMvVBeJJ9z84PQ7hDSQ05a2zjGbk6285fHmj5rs4D7gWLkagQrhBBGSZIUQIoQ4o6pePxXF2tlGzeS8+xzuHXuTLM5X6NwvvImJoTgzJ48EhemgAS9J0fRIsaH8l256PbkIarMuHVsjNeQUBw87Q69f3ZMBgP7Vy1h/4rFIEn4DOhARTtfcvV5NYIsvyK/tqlSUhKoCpRFGI4EndtKkF5Ls65PEtTxMbxcvK1uqiwpNzBpzh5yNXo+nhTHljOFLDqYhZuTkhnxETwSH4678y1OZlt4RvZnO78ZfJrLptGooX8tAWLUQ86hS+Is6wAYtPI571AI7QEh3eVH3xZ/rb+NHTv1wJpi7S1gnhAio45zrYQQp298mdblryzWdImJZD3xJK4xMYR8NxeF+5VJQPVaA1t/OkNaUjGBkd70G98cjhej258PZguubf3w6NsMp0BVHTPYuZupMFbUiqjM1mZTcioFr52FuOogrWk5B1qqqXA1A+Dn6kewKpggD9k8efnuWIBbAA5IsPsz2Pw2NAqFhO+haTubrF1baeTeb/eRXKBlwfQudG8umx3PFer4cP1Z1p3Mx0/lxNP9I5nSJQSnBvro3TQpG2XRVpwsFwgf8m9o0ubWrqEOjEYj2dnZVFZWXr9xfREWOb2JqQrMVbJp+GKBGwcneXfRwUV+VNgrQdixcxEXFxeCg4NxdKzt+2tNsdYNOCmE0FY/9wRaCSH23fiybcNfVaxVHDhA5iMzcIqIIPT7BTh4XhlBlX6smC0/nqZKb6LHwBCamcxUHCkEAW7tA/DoG4yj/y32AbJjddJL0zlYcLCWmTJHl0NJZUlNG1WFku5n/AjKd8bo7YhiUEuCY9rW31RZUQLLZ0LKemg9FkZ/brOoPb3BzAPz93M4Q80393ekf8vGV7Q5nKnmP2vPsC+thBAfN/42OIpR7QIb7K93U5iNcHC+nO6jshQ6PAD9XgfVVQrZ3wLS0tLw8PDA19f3xnc6zQY531yVTn406atPSHJRdCcVOKvA0d1epsuOnasghODChQtotVrCw2v7Z1pTrB0BOlyMAK02fx4UQnS48aXbhr+iWNMfO0bm9IdQNmlC6I8/oPSpnbzTUGli15JznNqZS7OmbnRupsKcogYHCffOTfDoHYyykb081N1Otjabr5K+4vfU37EIC0pJSVNV05rdsSBVEIEuTTDsOUf6hu0oFA50nzCFDsNH1y/Ks2aig3IRdm2+vIPUZYbNTFsGk4UZPxxkR0oRn01uz6jYwKu2FUKwLbmI99ed5XReGa2bevLS0Gj6RPnf2ujRihI5oe6Bb8HRDXq/AF1n3pYcYadPn6Zly5YNq+JgroKqcrlSgEEnizWQU2o4uoOzuyzQHN3siWnt2GkAQgjOnDlDq1a1o8itmbpDujxVhxDCIkmS/SvUHUDl2bNkzngUh0aNCJn33RVCLe98KZsWnEKhrmRQuAo3dRWW9FJU8cF4xAfh4GF3iL7bKaoo4ptj37AkZQkOkgPTWk1jUvQkAlWBtaIqUw8fYMu3cygtyCe6ezx96pPY9nKEgH1fw4Y35SLZD6+HoI42uCIZk9nCswuPsD25iP/c0/aaQg3kSNh+0QH0ifRnVVIu/914lgfnH6BbhA+vDGtFXDNvm621Fm4+MOw9OYHuhjdg49/lQITB70DLkbfcZ+uaQk0I2efMUL1rZtCBxSSfUyjByR3c/eVHR1dZsNmxY+eGuNkvjfURXamSJD0DfFX9/Akg9aZmtXPTVKWlkfnQwyhcXAhZMB/HxpfMQ2aThQO/p5G+KZN2KiW+KiWS3oxqQAiqHoE4uNuzsN/taCo1zDs5j/+d/h8mi4nxkeN5tN2jNHavbSasFeUZ1IwJb7xTE+VZb/QaWPUUnP4NoofD2C/BtZEVr6Y2FovglWXHWXM8nzdHtmZS55B691UoJMa2D2J426b8si+Dz7ecY+zsXQxr04QXhkTT3P8W+WP6RcK9C+HcZlj/OiycCmHx8m6kjXz7rovFAsaKS7tmhgoQso8iDk7g7CHvmjmp5J1AezCAHTt3DPX5qjQT6AHkANlAV+DmCkzauSkM2TlkTn8IhCBk/jycgoNrzl3I0bHp3f247syml0qJn7sjnkPDaPpKZ7wGhdqF2l1OubGcr5O+ZtiyYSw4sYBBoYNYNXYVb3Z/s5ZQMxkM7FnyPxY8/wSZx5Pofd907n//s4YLtdyj8E0fOLNG3h2a/ItNhZoQgrd/P8WSQ9k8OzCSh3vdWP4tJ6WCB3uGs/2lfjw7MJIdyUUM/ngHry47Rn6pFR3ur0eLATBzJ4z4r1wNYU5vWPkUaAtsP7ewQGWZnIS2OBnyj8GFFNDmyT52ro3kiM2AGDmNRqMwcPcDRxerCbX8/HwmT55M8+bN6dixI8OHDyc5Ofmq7dPT02nTRg7O2LZtGyNHjrzm+EePHmXNmjVWWetFVKorBb1Go+HLL7+84THDwsIoLi6+mWXxyCOPcOrUKQD+/e9/1xy//G9mTeq6ZpPZgqbCQFZJBf949z+0bNWa1jExTJ48+aqBLIsWLaJ169bExMRw77331jpXVlZGcHAwTz31VJ1933zzTdq1a0dcXByDBw8mNzcXgJ9//pl27drRtm1bevToQVJSUp39v/jiC1q0aIEkSVf8/bdt20ZcXBwxMTH06dOnzv6bN2+mQ4cOxMXF0atXL86dO3dFm40bN9KxY0fatm1Lx44d2bJlS51j3TQ1OaX+BD8dO3YUf3YMBQUiZdBgcaZzF6E/fbrmuNloFqd+Pi2Ov7BdZL28Q2S+tVtod2YLi8F0G1drx1rojXqx4MQCEf+/eNFmQRsxa8sskVKSUmfb84f2i2+fflh8OHGE+O3j90RZcVHDJ7RYhNg/V4i3/YT4byshMvbe5BXUj/+uPyNCX/5d/Ou3k8JisVht3CJtpfjHyhOixWurRdTra8T/rTktNOUGq41fLyrUQqx7TYi3fIV4N1CIHf8VwqC33vhl+UKcWCbE6heF+KqnOLV7vRA5h4XIOSJE4RkhNNlCVGiEMBmtN+c1sFgsolu3buKrr76qOXb06FGxY8eOq/ZJS0sTMTExQgghtm7dKkaMGHHNOebPny+efPLJBq3LaLz29bu7u19zXTdCaGioKCq6gffhVbh8jTe7tqtxcdzyKqPIL9WLlAKtOJalFklZarH54CkR1CxE7EvOFUlZajF45FjxzkdfivOFWpGjrhDFukqhqzSK02fOiLi4OFFSUiKEEKKgoKDWHM8884yYMmXKVf+HpaWlNb9/+umn4rHHHhNCCLFr166aMdesWSO6dOlSZ//Dhw+LtLS0K/7+arVatGrVSmRkZNS5rotERkaKU6dOCSGEmD17tnjggQfqnCMnJ0cIIcTx48dFYGBgnWMJIWrGuhzkGIDr6pvrmkElSXIBHgZigBpPdCHEQ7aRj3auhkmtJvOhhzAXFxMy7ztcWrZEmAUlu3MoWZ+Bh8lClbMD7oND8e4RiNSAFAZmk4l9yxeRnnSIyC49iOk7EDdPG9ddtHNdjBYjy1OWM+fYHAorCukR2IOn2z9NG7/a36QtZjPnD+7j6IbfyTxxDJ+gZiS8+S4hbWIbPmmVFn57Fk4sgRYDYdw3tyRL/5zt5/lsyzkmd27G6yNaWTUwwE/lzD9Hx/Bwr3A+2pjMnB3n+WVfBk/0a8GDPcJwuRWJn1295VxsnR6Sff82vwWH5sOgt+Wo2oZcrxBQklpdFaA6x1lJtXeKoxsEd5IjdH1bgKMbb60+w6ncTKteTutAT/4xKuaq57du3YqjoyMzZ86sORYbG1u9fMFLL73E2rVrkSSJN954g0mTJl11rP379zNr1iwqKytxdXVl/vz5hIeH8/e//x29Xs/OnTt59dVXGTlyJE8//TQnTpzAaDTyz3/+kzFjxrBgwQKWLVuGTqfDbDazevVqxowZg1qtxmg08s477zBmzJirzv/KK69w/vx54uLiGDRoEP/4xz/q7F9eXs7EiRPJzs7GbDbz5ptv1rouvV7P+PHjGT9+PDNmzKg5vnjxYvbs2cNHH33Ep59+yqeffkpqaiqpqalMmzaNXbt20bdvXz788EOWLFmCXq+v2RV69913MZvNzJgxg927dxMUFMTKlStxdXWtdQ0PPvggnp6eHDx4kPz8fN5//30mTJgAwAcffMCiRYuoqqpizJix/O3VN3jm+Rc5d/48Hdq3p1t8X954+9/4e7rg6azEBw8kYSHU2xEnV2eEsYqwkCAsQs6JaKl2c//4k9mMv+8hSs1OVJbqcVF5ozeYcFY6cOTIYQoKChg6dChXCwz0vCyzQXl5ec1nQo8ePWqOd+vWjezs7Dr7t2/fvs7jv/zyC+PHjyckRHaxCAiou8SZJEmUlZUBUFpaSmDglb6zl88RExODXq+nqqoK5zrynN4M9fFZ+xE4AwwB3gbuA+qVW02SJG/kRLptkJPxPAScBRYCYUA6MFHUUSRekqQHgDeqn74jhPi+PnP+WTFrtWQ9/AjGrGyaffMNLm3aoduXR8mGDBTlRowWgejQmBYTIlE0MM/UhexM1s7+iILUczQKDGbHz/PZtegnorv1InbwcJpGNiCizI5VMFvMrElbw5dHvyRbl02cfxzvxb9H5yada7XTlhRzfPN6jm9ej05dgoefP32mPUz7oSMbFuV5kYKTsOh++cY/4O/Q8zlQ2N6x/Od9Gfzf2jOMbNeUd8e1tdnrrZmPGx9PimNGfATvrz/De2vPsGBXOs8OjGRCx2CUDrfAid63OUz5RS4Ov+41Obo2pDsM/T8IrPvmgsUs/28ulmzK3Au6fPmcayO5f8fpcvLZprHg4AinT8t+aLeJEydO0LFj3UEoy5Yt4+jRoyQlJVFcXEznzp3p3bv3Vcdq2bIliYmJKJVKNm3axGuvvcbSpUt5++23OXjwIF988QUAr732Gv3792fevHloNBq6dOnCwIEDATh8+DDHjh3Dx8cHk8nE8uXL8fT0pLi4mG7dujF69Oirvu7ee+89Tpw4wdGjRwGu2n/dunUEBgayevVqQL7BX0Sn0zF58mTuv/9+7r///lrjx8fH8/777wOQmJiIr68vOTk5JCYmXvF3ee+99/jiiy9q1pKenk5KSgr/+9//+Pbbb5k4cSJLly5l6tSpV1xHXl4eO3fu5MyZM4wePZoJEyawfv16Tp85y2+bdlCmNzJjagIRazby9Mtvcvb0SfYcOISHs7LWe8MtOJgXXniB5uFhuLq6MnjwYO67ZzQgmy7jOnRk0NAR5Gel4ezowPjhAzCZzDz+3Mv07DcQYbHw+NPP8tmc79i3czsGk5lKoxlnpYIZM2Ywc+ZMOnWSAyRff/11fvjhB7y8vNi6desV1/Tdd98xbNiwmufDhw9n7ty5dQqriyQnJ2M0Gunbty9arZZZs2bV/E8u7z937lyGDx+Oq6srnp6e7N27F4BVq1Zx8OBB3n777VrjLl26lA4dOlhdqEH9xFoLIUSCJEljhBDfS5L0C5BYz/E/BdYJISZIkuQEuAGvAZuFEO9JkvQK8Arw8uWdJEnyAf6BXIdUAIckSVpVl6j7K2CpqCDrsZlUpqQQ/MnnWIxB5P3nABatgVKThSJvFzo82gbvgCsT4V5zXIuZQ6tXsmvhjzi5uDLq+VeJ6tqTosx0kjau5XTiFk4lbsU/NJzYQcNpFd8XJxfX6w9s54YRQrAlawtfHPmCc5pztPRpyewBs4kPiq+5kQiLhYwTSSRtWMP5Q/sQQhAe24GBM54kvH0nFDeaUuHIT7D6b+DiBfevgvB4K17Z1Vl5NIc3VpxgQMsAPp4Uh8MtyI/WOtCTBdO7sDf1Au+tPcMry47zbWIqLw6JZkhMk1vz5SSiL8xMhMM/wJZ34Jt+EHcv9H9TjirNOSzvmGXsgaz9UFV94/cMlv83FysD+EVfV1BfawfsdrBz506mTJmCg4MDjRs3pk+fPhw4cIB27eoOvigtLeWBBx4gJSUFSZIwGo11ttuwYQOrVq3iww8/BKCyspLMTHlHcdCgQfhUR8wLIXjttdfYsWMHCoWCnJwcCgoKaNKkSb3Wf7X+bdu25W9/+xsvv/wyI0eOJD7+0ntozJgxvPTSS9x3331XjNekSRN0Oh1arZasrCzuvfdeduzYQWJiIuPHj7/uesLDw4mLk/1RO3bsSHp6ep3txo4di0KhICq6JQUFBWSVVPDr8t/ZsGEDu3t0QZIkKivKqSzOoUXH1jg6KGjkdmXWALVazcqVK0lLS8Pb25uEhAR++uknpk6dyr/+9a+adg4ICrLS2bcrkcysLPr06cOufYf4+X8/M2DQEBr5N6VMb6Ss0kRygRZJknjp3Y9xcXSgUFuJi9KBf7z1Nu+8806NSH3rrbdqxt+6dSvfffcdO3furDlWHx9Gk8nEoUOH2Lx5M3q9nu7du9OtWzeioqJq9f/4449Zs2YNXbt25YMPPuD5559n7ty5jB49mtGjR9ca8+TJk7z88sts2LDhuvPfCPURaxffFRpJktoA+UDde4aXIUmSF9AbeBBACGEADJIkjQH6Vjf7HtjGH8Qa8i7eRiFESfVYG5Frkv6vHuv9U2GpqiL7qafQnziD31OfoN3thKU8FbWAMxVmwoeF0ndIKIoG7ghoCvJZ9+XH5Jw5SfNO3Rg040ncvWXHcf+QMAY+/Di9732A0zu3k7RxDZvmzmbHz/NoFd+fuEHD8AsJs8HV/nURQrAnbw+fH/6cExdOEOYZxod9PmRQ6CAU1SkT9DotJ7dt4timtajzcnH18KTTyHG0GzgM78b1u8nUiaEcVr8ASb/IGfjvmQuq677FrcLGUwU8vyiJbuG+zL6vA463YmfrMrpF+LL8iR6sP1nAB+vPMPOnw8Q18+bloS1rKiXYFIWDXBC+zXhI/C/s/QpOLJODA8xVchu/aGgzTq6pGdodvOsfHXs7iYmJYcmSJVYZ680336Rfv34sX76c9PR0+vbtW2c7IQRLly4lOjq61vF9+/bhfllVl59//pmioiIOHTqEo6MjYWFhDar0cLX+UVFRHD58mDVr1vDGG28wYMAA/v73vwPQs2dP1q1bx7333lvnl4EePXowf/58oqOjiY+PZ968eezZs4f//ve/113P5Ts5Dg4O6PX6WueFEJgsFvRmiXOFOvQGE2aLQFtpROkg8cJLL/P0E4/X2j27muAD2LRpE+Hh4fj7y0mfx48fz+7du6/YzQsODqZr1644OjrSPCKC6KgoCnMyOHn0IImJifzv+7nodDoMBgNNfL15+e9vU2m0UF5lQlNxqcSdgyTRc8hYZtx3D0+/8BoujgrOnj7JI488wtq1a/H1bdh7NTg4GF9fX9zd3XF3d6d3794kJSURFRVV06aoqIikpCS6du0KwKRJkxg6dGid42VnZzNu3Dh++OEHmje3TQXO+oi1byRJaoRsklwFqIA369EvHCgC5lcXgT8EzAIaCyHyqtvkA1emJIcgIOuy59nVx/5SCKORnOdfwagOwGPUR1SlKajwUHBIa0L4uzLoyTj8Qxpm5hBCcGzTOrb/+B2SQsHQJ56jde/+dX54OLm6ETtoGO0GDiUv5QxJG9ZwYusGkjasJqhla2IHDSeya0+UjvYI05vhaOFRPjvyGQfyD9DUvSlv93ibUc1HoVQoEUKQl3KW/2fvzONrutY//OwzZJAICSKCSKgYIoPEECXGxpCooKippqKD9laLKD/c1qVV1NAqbanQG9pqcampEQ2JoYgIYowhZBIJmcczrN8fJzmSZhCtJOh5fM4nZ6+91tpr7Tg53/2ud73vuYP7uHo8DLWqANtWbenyyihaenb7+/c++SpsGw/JV6DHh9DDv9qCnR67nsK0rRG0a1yH9eM7VI/fWBlIkkT/dja81MaaHRHxrDh4jVHr/6DK/QfqAAAgAElEQVSHYwNm929NW9uqyc5QApM6Ot81j4lwbLUutpldF92rGvwFq4LevXszd+5cvv32W6ZO1QUQOH/+POnp6Xh5efHNN98wfvx4Hjx4QGhoKMuWLStXMKWnp9O4se4rYNOmTfry2rVrk5mZqT/u168fX375JV9++SWSJHH27Nky/ZbS09OxtrZGqVQSEhLC7dulsimW4M/XKa99QkICVlZWjB07lrp167JhwwZ9m4ULF7Jw4UKmTZtW5s5SLy8vFixYwIIFC2jfvj0hISGYmppSp05p32GlUolKpSqVuqg4ao2WrHw1mXmFr1w16bkqQNDAwgSZBG0aWTDcbyDz589nysTxmJubEx8fj1KpLDXn4tjZ2fHHH3+Qk5ODqakphw4d0i9bFmfw4MH88MMPTJw4kZSUFK5du0bz5s3ZsmWLvs6mTZsIDw9n1efLSo5fq+Xi5as0adacPLWWbb/tpVnzliSk55IYH8uUV4fy6epvkFvaEp+Wi4lCholSjolShvwRlmY/Pz/eeecd1Go1BQUFnDx5kvfff79EHUtLS9LT07l27RqOjo4cPHiwVEBb0O2a9fX1ZcmSJXTt2rXC6/4dKhRrhdkKMgqXH0OB5o/ZtzvwrhDipCRJq9EteeoRQghJkipOofAIJEmaSmEokSJnwecBdWoudz/5EUx9MW5lhNSsLuGx2cTFZuPauymeg5ujMHq8L7fMBykEffMlMZFnsGvnSr+33sOi/qMtKJIkYevYBlvHNvQYN7nQunOAfV8ux3Tzetr18sb1pf7Usf4b1p1/IFceXOHLs18SGhdKPZN6zOk0h2GOwzCSG6HKy+P8sWDOBe3nXswNlCamOPV8CVfvATRo9tfCWZTi/DbdRgKlKby2A1r0fjL9VoIzt1OZ8n04zeubsXliR8yrOwl7GSjkMkZ0bMogN1s2H49h7eEb+H4Zhp+rLR94t8KuXjWkY7NygJdXVf11qgFJkti5cyfTp0/ns88+w8TEBHt7e1atWkW3bt04ceIErq6uSJLE0qVLsbGxKdea4+/vz/jx41m0aBG+vr768l69erFkyRLc3NyYM2cO8+fPZ/r06bi4uKDVanFwcGDPnj2l+hszZgwvv/wyzs7OdOjQgdatW1c4l3r16tG1a1fatWvHgAEDmD17dpntL1y4wKxZs5DJZCiVStatW1ein9WrVzNp0iT8/f31PmpFeHl5ERsbS/fu3ZHL5TRt2rTccU2dOhUXFxfc3d1ZvHgxoHsQz1VpyMxTkZpdwOXEDASgkEmYmygxM1bQxNKUF6wfPuBLkkTfvn25fPkyXbp0AXShSwIDA2nRokWJOS9b9lBMde7cmWHDhuHu7o5CoaB9+/Z6Qb5gwQI6dOjAoEGD6NevH0FBQbRt2xa5XM6yZcseaQWbPHmy3mdt4YJ5XL16FZlMRrNmzdj4zToaNrLg8/kryUhP5ZN5M9AKkMnl/LBX5882bdxwFn2+BrumTQj87mvWfbGSe0l3cXFx0fujtWnThv79++Pi4oJMJmPy5Mn68CfFfdbWr1/PK6+8gkwmw9LSko0bNwIlfdbWrFnD9evX9WIcdMvx5W1a+KtUJt1UuKhEKoQy2tkAfwgh7AuPvdCJtReAnkKIREmSGgGHhRCt/tR2VGGdNwqPvymsV+Ey6POQbkr9II+Mw7HknEpAaAVyiyySWjnyR0g8teoY0Xt8G5q2tnp0R8UQQnDl2BEObVyHRqWm+9iJuHn7IP0Nx3Gh1XL7/FkiD+7n5plTCAQObh64evvg0N7jr/tN/QO4lX6LtZFrORBzgNpGtZnUbhKjW4+mlrIW9+PuEBm0j0uhv1OQm0MDO3tc+/rQpltPjEyfkFhQ5cGB2XBmk25pbdh3YFFxhoAnyaWEDEZ+ewIrMyO2vdkF69pPZ7qz9FwVXx+5QcCxW2i0gtGd7Hi3T0vqm1d/6qi/wuXLl8u0BBh4fviz9Uyt1S0d1jKSY26ixMJYgamR/LnfICaEQKXRkqfSkqfW6H6qNOSrtRRpHAkwUugsb8aFP02UcowUMmTVdH/K+kw+ydygS4AUdDs4s4vKi/zJHtE2DJgshLgqSdJHQJHTwP1iGwyshBD+f2pnhW7ZtCj/aATg8ahrPstiTXUvh8yQWHLO3QOtloJbYYj2NpyV3Ei6lUHLjg3pPtIRk8cMapuTkU7whq+IPnmcRi1b0f/tD7CyfbIryhkpyVz4XbcjMTstFYsG1rj06U+7Xt56PzgDkJCVwNfnvmbXjV0Yy415re1rjHcaj5nMlOiTxzl3cD9xl6OQKxQ4enbD1dsH21ZPNoQF92/Az+Ph7gXo9j70mletCbhvJGfx6jcnMJLL2PZmF5pYVoO16m+SlJHHquBotoXH6nareTVnSvfmT4U1sCIMYu3546H1TCfOcgvUJaxntU0UpXZu/pMRQpCv1pKv0pCn1gm4PJWWArWGIuUjSRLGRUuoxZZSlXLZExe5VS3WbpVRLIQQj1wSlSTJDV3oDiN0KaomosuasA2wA26jC93xQJKkDsCbQojJhW0nods5CrBYCBHwqOs9i2KtID6LzJA75F68j6SQIRklkr79c1J9JnMhqwVyhYweo1vRskNZrn0Vcz38JAe//ZK8rCxeHDGGjoOGVqnFS6NWcyP8D84d3MedqPPI5ApaduqCa18fmrRp99w/3ZVHSm4KGy5sYNvVbUhIvNr6VV5v9zrKLA3nDx3gwu9B5KSnUce6IS4vDaBdL++qiXF38X+66PkyOQz9Fhz7PflrVEBcag7Dvz6BSqNl2xtdaF5dqZ+eEDeSs/g86Cr7LtylnpkR7/R+gdGd7TBWPJ1WZINYez4o33qm0Iuzf4L17Emi1QryiyxwhT/zVRoKNMU2Ncgk2jSyeKJWtyoVa88Sz5JYy49JJzMklryrqUjGcsxftEUVE0LcV19zvedMkgrq0bStFb1fa4O55eMtu+TnZBOyaT0XjwTToJkDA6Z98OT8nCrJg4Q4zh3cz8UjweRnZ1OviR2u3gNo2703xrUeL8TIs0p6fjqbLm5iy+UtFGgKGPzCYKY6TyXvejzngvZx6+wZAJp7dMTV2wd7l/Z/a2m6XNT5uiCsp76BJh1hWADUbfrkr1MB9zLzGPH1CR5kF/Dj1C7V47RfRZyLTeOzA1c4fuM+TSxNmdHXET/XxsiqIeTI42AQa88mButZzaHRPlxKVWsEDS2erItGVVvWxpVVLoT4/nEGWR087WJNCEH+9TQyfo+l4FY6MjMF5t0aY97FlrQd27iw9leutRuPRmlC11deoF2Pxo/9tHQn6hwH1q0i6/59Og0eTpdhI/9acNQnhCo/jyvHQzkXtJ+km9EojI1p060nrt4+NHSomi3ONU2OKofAy4FsitpEliqLAQ4DmNT8NdLCL3M++AAZyUmY1bXEuXdfnPv0q9Qmj79M6m1d0NWECPCcBi99BIrScZOqkrScAl795g9iU3MInNwZd7tnf2lcCEFYdAqfHbjCxYQMWtvUZnb/1vRs1eCpsXAYxNqzg8F69s+gqsXal8UOTYA+QIQQYthfGGuV8rSKNaEV5F1+QMbhWFSxmcgsjKjdvQlmnWyQGcm5t+1/hG29xF0bTxrYmeM9yQlLm8ezPqny8wjbupmzB37FslFjBkz7gEYtWz26YTVy90Y05w7u48qxUNQF+TR6oRWufX1w7NINpdGz4bRdEfmafH6++jPrL6znQd4Dejbpwejavjz4I4rok8fQqNU0beuMa19fXujoiVxRxT5PV/bB/97UhZUe/BW0eblqr1cGWflqxmw4yeXEDDZN6MiLL9Sv9jFUJVqtYM+FRD4Pusrt+zl0crDiwwGtnwpBahBrTy8G69k/k2pdBi1MIfWjEKLs6HA1yNMm1oRWkHshmYzfY1En5SC3MqF2jyaYeTTU5+289t8gQg9lUGBcF/f+zeg4qAXyx/yAJly7woG1K0hNTKD9gJfxGjUepfHTucMOIC8ri0uhh4g8uJ/UhDhMzGvj1KMPrt4DsGz07IXTU2vV7Lq+i3Xn1pGUk0SXeh3xy/ck5cQ5UmJvY1zLjLY9euP6kg/1mlTD8qNGpcs7efxLXeqh4Zt1ISGqmTyVhgkBpzgdk8o3Yz14qe3j+10+KxSotfx4+g5fHIomJauAfk4NmdWvVYkwCdWNQaw9XRisZwb+jlj7K7I9G13AWwPlINRask/fJWnFGR78cBWEwPLVVtjM6IB550ZIChlqlYaQVSEcPCpDoZQzZLoznkNaPpZQ06hVhP2wmR8X+KNWqRg+fzG9J7zxVAs1ABNzc9x9/Ji4Yh3D53+CXTtXzh74lY3T3+CXxfOJPnUcrUZT08N8JFqhZd/NfQzeNZiPTnyEfX59PkwbQrttGVzZthu5UknfN/7FG+s203vCG9Uj1FJvwyZfnVDrOBkmBdWIUCtQa3l7SwQnbz1gxQjX51qoARgpZIzrYs+RWb34wNuRY9fv03dlKLN/OU9CWu6jO3hOuXv3LiNHjqRFixZ4eHjg4+PDtWvXyq0fExOjj3d1+PBhBg4cWGH/kZGRlUov9DiYm5fe+JKWllZmINuKEEKQU6AmKSOPJnbNOH7xFnce5JCZp8LcREFTq1q0bWTBC9bmNLQwoZaxokKhNnnyZC5dugTAJ598oi8vfs+eJI+a88qVK3FycqJdu3aMGjWqzIDGX3/9Nc7Ozri5udGtWzf9+A8ePIiHhwfOzs54eHjw+++/l3mNn3/+GScnJ2QyWalk7+fPn6dLly44OTnh7Oxc5vUjIyPx9PTEzc2NDh06cOrUqTLrFPXj4uLCTz/9VOF9qTGEEBW+gF/RZS7YDexBt6tzyaPa1cTLw8ND1CTaArXIPBonEj45KWJnh4q7X0SInAvJQqvRlqiXHJshAmeHiDVvHBK7x60WuSmpj32tezE3xeaZ08TyEb5i/9qVIi8760lNo0bIfHBfnPjlB/HNWxPE8hG+4us3x4lj27aIzPspNT20Umi1WhFyJ0QM3TVUuH7XTkxdOUismzVFLB/hK1aNGSL2r10pEqOvVt+A8jKEOLtFiE0vC/HvOkIsthXi/M/Vd/0/odZoxdtbzohms/eILX/crrFx1CQpmXni490XRcu5+0TL/9snFu+9JFKz86t1DJcuXarW6/0ZrVYrPD09xbp16/RlkZGRIjQ0tNw2t27dEk5OTkIIIUJCQoSvr2+F1wgICBDTpk17rHGpVKoKz5uZmVU4riK0Wq1QazQir0AtsvJUIjU7XyRn5onEtFxx+362uBifLs7FpopzsamicVM7cfFmrMjOUwmtVluq/8el+BjLGtuToKJ+4+LihL29vcjJyRFCCDF8+HAREBBQql56err+/a5du0S/fv2EEEJERESI+Ph4IYQQFy5cELa2tmVe59KlS+LKlSuiR48e4vTp0/pylUolnJ2dRWRkpBBCiJSUFKFWq0u19/b2Fvv27RNCCLF3717Ro0ePUnWuXr0qrl27JoQQIj4+XtjY2IjU1Mf/Tq4MZX0mgXBRCX1TGaeZ5cXeq4HbQoi4J6wZn2m0eWqy/kgk62g82iwVRvYWWA59AWNHyxJPSlqtIPLgHU7uuoEiLwOPjGA6rP8PCsu6lb+WRsPp3ds5/vNWTMzNGew/nxYenatiWtWKuaUVnq+MpNPg4dw8G865oL2c+GUrf+z4kRc6eOLq7YNdO5eq2S35GJxMPMkXEV9w6/YlOiY1xutOKzQ5eRg1kug4bgpOPfpgUsaT+RNHo4Ybv8P5n+DKXlDngqW9Ll1U+7E1lj9SCMHcHRfYez6R//Npw+jOz09WkcehnrkxC15uy8Su9qwMvsb6sJv8cOoOb/VswcQXHTB9zOwjf5v9H+pi6z1JbJxhwJJyT4eEhKBUKnnzzTf1Za6uroDu/4m/vz/79+9HkiTmzZvHq6++Wm5fp06d4r333iMvLw9TU1MCAgJwcHBgwYIF5ObmcvToUebMmcPAgQN59913iYqKQqVS8dFHH+Hn58emTZvYsWMHWVlZaDQa9u7di5+fH6mpqahUKhYtWoSfn58+gGqeSoNaK1BrtKi1gukzZnH9xg3atnPhxR69eOv92bw9YRQZ6WmoVSremTWPXv18yMnJxv+tidy7m4jQavCf83+MHzMKhUzCurYJklaFj8/LDB06lClTpujn9/PPP3PixAlWrFjB6tWrWb16NTdv3uTmzZu89tprHDt2jJ49e7J8+XJ++eUXcnNzcXNzw8nJicWLF6PRaJgyZQrHjx+ncePG7Nq1C1NT0xL3cMKECVhYWBAeHs7du3dZunQpw4bpXM+XLVvGtm3byM/PZ8iQIXz88cd8+OGH3LhxAzc3N7y9vUtkMABdMvTc3FyUSiU5OTnY2pYOrG1h8XDXd3Z2tv77sHgKMCcnJ3Jzc8nPzy+R4xQodxk/KCgIFxcX/f+n8rIiSJJERkYGoEsRVtYYi+cDtbW1xdramuTkZOrWrfz3cnVQGbF2B0gUQuQBSJJkKkmSvRAipkpH9gygzVGReSyBrGMJiDw1xi3rYtHLDuPmpWNkZaTkErzpEonX07FOvYBTeggtv/8WhWXlHZEfJMRzYO0KEqOv4ujZjT6vv1U18bhqEJlczgsdOvNCh86k3U3kXPB+og4HE33qOJaNbHH19qFtjz6YmlevL9D55PN8cWY1Cecv4BJXH/ekxkgyGc07uhcKSdeq9zURAhLO6gRa1HbITgaTuuA2ClxGQtNOUIP+LkIIFu29zE/hsfyr9wtM6f442emeT5pa1WLFCDemdm/OsgNXWXrgKpuPx/BeH0dGdGjyXDuQR0VF4eHhUea5HTt2EBkZyblz50hJSaFjx45079693L5at25NWFgYCoWC4OBg5s6dy/bt21m4cCHh4eGsWbMGgLlz59K7d282btxIWloanTp1ok+fPmi0Ws5ERHDidAQWdeqSVaDi600/YGpuzr17yQzp3wvHjj1RC9AKuJZUMifmmzPnc/HiRf536BgKmYQktPz3h21Y1q1Leup9+vbqzpSxI/h190EcHewIPRQE6ARC0e84KyuLkSNHMm7cOMaNKxlkwcvLS59+KiwsjHr16hEfH09YWFip+7JkyRLWrFlDZGQkoFsGjY6O5ocffmD9+vWMGDGC7du3l0qqDpCYmMjRo0e5cuUKgwYNYtiwYQQFBREdHc2pU6cQQjBo0CBCQ0NZsmQJUVFR+usUp3HjxsycORM7OztMTU3p27cvffv2BUqmmwL46quvWLFiBQUFBWUud27fvh13d3e9UCuebqo8rl27hiRJ9OvXj+TkZEaOHIm/v3+p9qtWraJfv37MnDkTrVbL8ePHAQgPD+frr78ukbsVdA8FBQUFVZaM/e9QGbH2M/BisWNNYVnHKhnRM4Ams4DMsHiy/0hEFGgwaVsPi15NMWpaWkAIIbhyIpGwn6JBaHG68wu2uVdwCPwvykrmDhNaLWd/20vY1k0olEp8/jWL1i92f+4dUevaNKLH2El0HTGWayePcS5oH4e/38DRH76n1Yvdce07AJsWjlV6H64+uMq6Y6tIPnmBNnEWtMm1xszKCtfhA3Du3Rdzq2pIsp16Gy5s0+XyTLkGciNw7A8ur0LLvtUeiqM8Vh+K5rujt5jY1Z73vR0f3eAfRGsbC76b0JHTMQ9Ysv8Kc3deYEPYTWb2a8WAdjZV/1muwAJWExw9epRRo0Yhl8tp2LAhPXr04PTp07i4uJRZPz09nfHjxxMdHY0kSahUKtRaLSqN7pWWU4BaK9i7/ze27/wfnyxZigCysnM5FH6JhLQ8OrzYgzSNEWkPclCpVCz/+P+IOHUCmUxG0t1EctLvY2vbCEnSiWyFTEIhl6GQSdRW1cZYIcOxoe5vvEql4v335xIaGopMJiMxIZ6M1BTau7oye9YsZs+ezcCBA/Hy8tLPwc/PD39/f8aMGVNqfjY2NmRlZZGZmUlsbCyjR48mNDSUsLAwhg4d+sj76eDggJubGwAeHh7l5lgdPHgwMpmMtm3bkpSUBOisVEFBQXprV1ZWFtHR0RXm2k5NTWXXrl3cunWLunXrMnz4cAIDAxk7dqw+P2YR06ZNY9q0aWzdupVFixaxefNm/bmLFy8ye/ZsgoKC9GV/FlBloVarOXr0KKdPn6ZWrVr06dMHDw8P+vTpU6L9unXrWLlyJa+88grbtm3j9ddfJzg4mA4dOpS6TmJiIq+99hqbN29GVsMrOGVRGbGmEEIUFB0IIQokSXo6vh2qGXVqHpmhcWSfvgsagalrAyx6NkVZTpiNnIwCDm+5wq1zKTRqZsoLwUswyX9As/9+j7Jx5XY9ZqTc47d1q7gTdR57Nw/6vfGv6hEITxEKIyPaevWirVcvkm/f4tzBfVwKO8zFI8FYO7TA1duHNl17oDR5chsrYtJjWL93GRknL2OXZEpTUZemzq607zeQFu6dkMmreBkrN1WXceD8NrijexrE7kV4eRq09QPTmg8NUZwNYTdZFRzNcI8mzPdt+9w/SPxVOtpb8cubXQi+fI+lB67w9pYIXJvUYXb/1s9dWBMnJyd++eWXStcXQqDWahFCF/IlK09FgVrL3fQ8PvCfg0vHF1n2zffcjolh3Cu+XErI4G56Hum5Ku48yAFApdGw4tv/0tLREYVMQimXoZBL3LlyDmsrCxzqm6GQy9j6381oczOIOncWpVKJvb099UxlNKpjigRY1ir5Fffn/89btmwhOTmZM2fO6Nvn5eXh6OhIREQE+/btY968efTp04cFCxYA0LVrVw4cOMDo0aPL/Hy8+OKLBAQE0KpVK7y8vNi4cSMnTpzg888/f+S9K758KJfLyc0te1NL8XpFS75CCObMmcMbb7xRom55gg8gODgYBwcHGjRoAMDQoUM5fvx4mda8IkaOHMlbb72lP46Li2PIkCF8//33j23JatKkCd27d6d+fd1nxsfHh4iICPr06VOi3ubNm1m9ejUAw4cPZ/LkyWX2l5GRga+vL4sXL8bT0/OxxlJdVEY+JkuSNKjoQJIkP3S5Qv9RCK0gef0Fsk/dxax9Q2xmdKDeyNblCrVb51P48T8nuX3xPp79GuJ8aAHGWUnYffcdxg6P3p0nhCDqcDCbZ75DYvQ1vKe8w9APP/rHCbU/06CZAy9NnsYb6zbT5/W30Wo0HPz2S755azy/B3zD/bjYv9X/7Xs3WPzVW2x8/w2sdt3GPq0Orv19mbTqG0bMW0zLjl2qTqipC+DyHvhpLCx3hD3TdUudvefBe+dh0n7wmPDUCbUfT91h0d7L+Do3YskrLk9dJP+nDUmS8G7bkAPTu7NsmAvJmfmM3nCS1747SVR8eo2OrciZWasVaLRa1IWWqwK1Lp9ivkpDnkpDboGGnAI12flqvbDKzFORkasiPVdFWk4B7T27kZ2bx4ovv+J+Vj4pWfkcPhHO7t+Caefeic2BW7mRlMHJS7c4FHIEC7s2RCdlka/WcDM5i6TMfPLUGpIz80lNS6d+w0bIZTL2bP8RmSTRqI4pdjb1kKnzaNmwNm0aWeDnO4D9PwXQ0tqc5g3MSbl9lUZ1TKltosRYIae2iRJTpZyszEysra1RKpWEhIRw+/btCu9L7dq1ycx8uDSanp5eZvuEhARq1arF2LFjmTVrFhEREfo2CxcuxNLSkmnTppV5DS8vL5YvX0737t1p3749ISEhGBsbU6dOaVcXpVKJSqX6K7/iUvTr14+NGzeSlZUFQHx8PPfu3Ss15+LY2dnxxx9/kJOTgxCCQ4cOlelfFh0drX+/d+9eWrZsCeh2mvr6+rJkyRK6du36l8Z84cIFcnJyUKvVHDlyhLZt25aqZ2try5EjRwD4/fff9dcvTkFBAUOGDGHcuHF6H76nkcpY1t4EtkiStKbwOA4oM6vB84wkk7Aa1hK5lSmKuuUHcC3IU3Psl+tcOppAvSbmDJzYnMyZU1GnpmK3KQCTVo9eHspOS+Xg+jXcCD9Jkzbt6PfWdOo2tHmS03nmMa5VC7e+Prh6DyDh6mUig/ZyPng/Zw/8SpO27XD19qFlpy6Vzt5w9VI4u7evQ3vpLiZaCaNGVnQbPRL37v2rNmCvEBB7Cs7/CBd36ixqZg2gw+vgMgJs29eoH9qj+PVcAnN2XqBnqwasfNUNuUGoVRq5TGJ4h6a87GrLf0/c5qvD1xn45VFedrXF3a4uGq1ApdE5uauKOburNNoS54rK1Bqhc4rX6t5PdTFGeS8LEAihi42s+/nwGCGKlT+0tjwpPlu3mWUfz2Xl8uUYmZjQuElTZn30KR17duDY8RMM6OGJJJP4v48X0bp5UxJi76CUy3Cob0aipSm1jRW0a2zBfxbMZfz48Wxc8zm+vr7IZRINahvzcn9v1qxcTpeOHsyZM4cFCxYwffp0XFxc0Gq1ODg4sGfPnlLjGjNmDC+//DLOzs506NCB1q1bVziPevXq0bVrV9q1a8eAAQOYPXt2me0vXLjArFmzkMlkKJVK1q1bV6Kf1atXM2nSJPz9/fU+akV4eXkRGxtL9+7dkcvlNG3atNxxTZ06FRcXF9zd3Vm8ePHj/EpK0bdvXy5fvkyXLl0AXeiSwMBAWrRoUWLOxTcYdO7cmWHDhuHu7o5CoaB9+/ZMnToVKOmztmbNGoKDg1EqlVhaWuqXQNesWcP169dZuHChftk0KCgIa2vrEj5nO3fu5N133yU5ORlfX1/c3Nz47bffsLS05IMPPqBjx45IkoSPjw++vr5ASZ+19evX895776FWqzExMeHbb78FSvqsbdu2jdDQUO7fv8+mTZsA2LRpk35Z+Wmh0kFxJUkyBxBCZFXpiP4GNR0UN/FGOsEBF8m4n4d7Xzs8etQn7vVJFNy6hd36b6nV8dFuftdOHiN4/VcU5OXSbeQ4PHz8anwH5LNCTkY6USEHOR+8n/R7SdSqUxfn3v1weanslE6q/DwiQ4MI29zUrcYAACAASURBVPMj4m4GKrkW0cYav1feom3bKt5he/8GnPtR54uWGgMKU2jtC64joXlPkNdcirDK8vuVJKZ+fwb3ZpZsntip+nc4Pmdk5Kn49shNvjt6i1xV6TiDOh8qCaVMhlwuoZDJUMoflinkEvKiMpnE9A5mNHF4AUmSkHio+YsfS8WOkUBC0pUXnSs8LnFO3/bhMeX0KRV2XGa7p/ghxICBqqCq0019AiwVQqQVHlsCM4QQ8/76kKuGmhJrGrWW03tuEfHbbcytTHhpQltsbJXcmfQ6uZcu0XTtWsy9ulXYR15WFoc2ruPKsSM0bP4CA6Z9QL0m/8ywB38XodUScy6CyIP7uBWh+//g4N4BN28f7F3deZAYz5nffuXCkYOQpyLVvAB5+6aMeWUGjo1Km9KfGNkpELVDt5szPhyQoHkP3UaBNi+Dcc1Fu39cjt9IYULAaVrb1GbL5M7UNnn6xeWzQm6BbqlRXkyEKWTSY4sbQwYDAwaeLv6OWKvMMugAIcTcogMhRKokST7AUyfWaoIHCdkcDLhISmwWbV5sRLfhLVHINMS+8Sa5UVE0XrXykULtVuQZgr5eTU5GOl2GjabzkBFVnzeyDIQQ3M+7j1ZoC59+dV8ORe/1/4rK/1RW+Hxe5ntJ92heqn5VPF1LMhkO7Tvg0L4DGSn3OB/8Gxd+/40dZ05hamFBbkYGWkkQY5ODaYcWvD5gBm3qVdGXmioXru7XCbTrwaBVQ8N24P0fcB4GFqXj/jztnL2TypTN4djXq8XmiZ0MQu0JY2okN1gpDRgwUILKKAK5JEnGQoh80MVZA579rNt/E6EVnA+J48TOGxiZyhnwpjPN3RogCgqIe/c9ck6exPazJVh4e5fbR0FeLkf++x3ngw9Qr4kdg/0X0LD5C9U4i4dkFGQw/+h8fo8tO+1HVfNnQViRsCtVV1dYoaiU9ZZolKDAOjaFJJtszDxaMu3F6bhZV4FfglYLt4/qBNql3ZCfAbUbgefbOiuazZNPDVNdXE7MYELAaerXNibw9c5Ymv0jN4YbMGDAQLVSGbG2BTgkSVJA4fFE4PuqG9LTT+aDPA5tvkz81VTsXerTa2xralkYITQa4mfPJuvIEWw+/pg6gwaV20fc5SgOrFtF+r0kOrw8lK4jxqIwqpkvvqsPrvL+4fdJzEpkivMUbMx0mxmEEOj/Fb4vUV609fvP5RXUL/pHkaNz8fJi/ZVVt6xrlKpf1nWL6tqBrIuM4c1eokujLk/eqnfvcqEf2s+QEQ9G5rowGy4jwN4LZM+2teRWSjavfXcKU6WcwNc7Y23xdOegNWDAgIHnhUeKNSHEZ5IknQNeKiz6jxDit6od1tOJEILo00kc+eEaWq2g19jWtOnaCEmSEFotifMXkLn/ANb+/li+OqLMPtQFBRzbFkj4np3UsW7Iq//+lCZtas7SsjN6J4tPLqaOUR029t9Ie+v2j25k4CGZd+HCL7rdnHcvgCSHF/qA90Jo5QNGtWp6hE+E+LRcxm44iRCCwMmeNLV6PuZlwIABA88ClXKMEkIcAA4ASJLUTZKkr4QQZQeLeU7RarQcDLjE9fB72DSvw0sT21Cnge4LSwhB0iefkr5jB/XfeYd6kyaW2UfSzevs/2oF9+Pu4PJSf3q89jpGJqZl1q1q8tR5fHrqU3ZE76CzTWc+6/4Z9Uz/2THcKk1+FlzZo1vmvHkYhBZs3aH/Z9DuFTBvUNMjfKIkZ+YzdsNJMvJU/DDFkxesqyH3qQEDBgwY0FOpmBCSJLWXJGmpJEkxwH+AK5VsFyNJ0gVJkiIlSQovLPup8Diy8HzpxGPltK1JZHIZpuZGeA5uzpCZ7nqhBpC8chWpgYFYTZxI/Wlvl2qrUas5/vNWts6bQX52FkPnfIz3lHdqTKjFZsYybv84dkTvYIrzFL7x/sYg1B6FRq3bILB9CixvCTvfgPvXwWsGTDsNU0PA883nTqil56h47buT3E3PY9PEjrRr/HzlojVQtdy9e5eRI0fSokULPDw88PHx4dq1a+XWj4mJoV073UrD4cOHGThwYIX9R0ZGsm/fvic6ZnPz0g8jaWlprF279i/3aW9vT0rK34slP3nyZC5dugTAJ598oi8vfs+eJI+a88qVK3FycqJdu3aMGjWKvLy8UnVWrFhB27ZtcXFxoU+fPqWCD2dkZNCkSRPeeeedMq/x888/4+TkhEwmo3ikh4MHD+Lh4YGzszMeHh5l5hwFmDVrFq1bt8bFxYUhQ4aQlpZWmak/nRRFrP7zC3AE/o1OmB0F3gVul1e/nD5igPoVnP8cWPBX2pb18vDwENVN8tffiEutWouEBf8WWq221PmU2Nvivx++J5aP8BV7v1gmcjMzq32Mxfn99u+iy5YuosvWLuLwncM1OpanHq1WiPizQuyfI8SylkL820KIT5sKsftfQsQcE0KjqekRVilZeSox+KujouXcfSLsWnJND8fAY3Lp0qUavb5WqxWenp5i3bp1+rLIyEgRGhpabptbt24JJycnIYQQISEhwtfXt8JrBAQEiGnTpj3WuFQqVYXnzczMKhzXX6FZs2YiOfnJfYaKj/Hvjq08Kuo3Li5O2Nvbi5ycHCGEEMOHDxcBAQGl6v3+++8iOztbCCHE2rVrxYgRI0qc/9e//iVGjRpV7u/w0qVL4sqVK6JHjx7i9OnT+vKIiAgRHx8vhBDiwoULwtbWtsz2v/32m/737e/vL/z9/SuYcdVT1mcSCBeV0DcVLYNeAcKAgUKI6wCSJL3/pESipPPuHgH0flJ9VjcP/htI8sqVWAx6GZt/LyjhsC60Ws7s28XRH79HaWLKy+9/iKNnxSE8qhK1Vs2as2v4Luo72li1YUXPFTSp3aTGxvNUkxar2yRw/idIvgIyJTj2e5g4Xfn8O9bnqTRM+T6c83HprB3jTreWz1feyn8an536jCsPKrUgUmlaW7VmdqfZ5Z4PCQlBqVTy5ptv6stcXV0BnZHA39+f/fv3I0kS8+bN49VXXy23r1OnTvHee++Rl5eHqakpAQEBODg4sGDBAnJzczl69Chz5sxh4MCBvPvuu0RFRaFSqfjoo4/w8/Nj06ZN7Nixg6ysLDQaDXv37sXPz4/U1FRUKhWLFi3Cz8+v3Ot/+OGH3LhxAzc3N7y9vfn3v/9dZvvs7GxGjBhBXFwcGo2G+fPnl5hXbm4uQ4cOZejQoUyZMkVf/vPPP3PixAlWrFjB6tWrWb16NTdv3uTmzZu89tprHDt2jJ49e7J8+XJ++eUXcnNzcXNzw8nJicWLF6PRaJgyZQrHjx+ncePG7Nq1C1PTkis3EyZMwMLCgvDwcO7evcvSpUv16ZWWLVvGtm3byM/PZ8iQIXz88cel5lw8gwHokqnn5uaiVCrJycnB1rZ0GKJevXrp33t6ehIYGKg/PnPmDElJSfTv35/y4qOWFyewKOk86HLQ5ubmkp+fXyL3KeiyMxS//uPkqn3aqEisDQVGAiGSJB0AfgQed/ucAIIkSRLAN0KIb4ud8wKShBDRZTetsK0eSZKmAlNBl6+sukjbvp2kxYup7f0Stp98UiLLQFrSXX5bt4q4y1G06NAZ7ynvYFa35vI5puSm4B/qz+m7pxnmOIwPO32IsfwfH32lJHnpcGmXLnF6TJiurKkn+K4ApyFQy6pmx1eNqDRa3tkawfEb91n5qiv9nAypzgw8PlFRUXh4eJR5bseOHURGRnLu3DlSUlLo2LEj3bt3L7ev1q1bExYWhkKhIDg4mLlz57J9+3YWLlxIeHg4a9bosiHOnTuX3r17s3HjRtLS0ujUqRMvvaTbGxcREcH58+exsrJCrVazc+dOLCwsSElJwdPTk0GDBpW7Q3zJkiVERUURGanz2imv/YEDB7C1tWXv3r2ALodoEVlZWYwcOZJx48YxblzJjI1eXl769FNhYWHUq1eP+Ph4wsLCSt2XJUuWsGbNGv1YYmJiiI6O5ocffmD9+vWMGDGC7du3l5lUPTExkaNHj3LlyhUGDRrEsGHDCAoKIjo6mlOnTiGEYNCgQYSGhpaac3EaN27MzJkzsbOzw9TUlL59++qFUfF0U8X57rvvGDBgAABarZYZM2YQGBhIcHBwiXrF00VVhu3bt+Pu7q4XauW137hxY4UPBE875Yo1IcT/gP9JkmQG+AHTAWtJktYBO4UQQZXov5sQIl6SJGvgoCRJV4QQoYXnRgE//MW2xcf5LfAt6DIYVGJMf5uM/ftJnL8As27dsP38c6TCALZCCC4c+o3D329Aksno//b7tO3eu0bTqkQkRTDzyEwyCzJZ1HURfi+U//T4j0NdADcO6cJtXN0PmnywagG9/g+ch4OVQ02PsNrRaAUztp0j+PI9/jO4HUPaG6yvzwMVWcBqgqNHjzJq1CjkcjkNGzakR48enD59GhcXlzLrp6enM378eKKjo5Ekqdwk5kFBQezevZvly5cDkJeXx507dwDw9vbGykr30CWEYO7cuYSGhiKTyYiPjycpKQkbm8o9mJTX3tnZmRkzZjB79mwGDhyIl5eXvo2fnx/+/v6MGTOmVH82NjZkZWWRmZlJbGwso0ePJjQ0lLCwMIYOHfrI8Tg4OOhzWXp4eBATE1NmvcGDByOTyWjbti1JSUn6exYUFKS3VmVlZREdHV2h8SM1NZVdu3Zx69Yt6taty/DhwwkMDGTs2LH6XJ/FCQwMJDw8XJ9Ufe3atfj4+NCkSem/Lxs2bHjkfIu4ePEis2fPJijooRwpq/3ixYtRKBRl3vtnhcqE7sgGtgJbC1NNDQdmA48Ua0KI+MKf9yRJ2gl0AkIlSVKgs9yV/dhVQdtHzqiKyQwJIX6WP6bu7Wny5RfICmOjZT24T9A3X3Ar8gx27Vzp99Z7ZeajrC6EEHx/6XtWnllJY/PGrHtpHa2sWtXYeJ4ahIC4cN0SZ9R2yH0AteqBx3hwGQmN3Z/qxOlViRCCef+LYve5BD4c0JrXPJvV9JAMPMM4OTk9sWWn+fPn06tXL3bu3ElMTAw9e/Yss54Qgu3bt9OqVcm/dSdPnsTMzEx/vGXLFpKTkzlz5gxKpRJ7e/syHeTLo7z2jo6OREREsG/fPubNm0efPn1YsGABAF27duXAgQOMHj26zAf4F198kYCAAFq1aoWXlxcbN27kxIkTfP75548cT/HlP7lcTm5u7iPrCfEwHuWcOXN44403StQtT/ABBAcH4+DgQIMGug1VQ4cO5fjx42Va84KDg1m8eDFHjhzRX//EiROEhYWxdu1asrKyKCgowNzcnCVLljxyrkXExcUxZMgQvv/+e1q0aFFuvU2bNrFnzx4OHTr0TOejfawM4UKIVCHEt0KIPo+qK0mSmSRJtYveA32BqMLTLwFXhBBxf6FtjZF94gTx703HpE0bmn79NTJTU4QQXD52hM0zpxF7KYreE99g2P/9p0aFWmZBJh8c/oDl4cvp1bQXPw788Z8j1NT5kHob7vyhy8N5Yi0EzYNfXocAX1jpBN+9BGf/q0uYPuonmHEVfJZBE49/tFD7dP8Vfjh1h2m9WvBmj/L/+BkwUBl69+5Nfn4+33770IPl/PnzhIWF4eXlxU8//YRGoyE5OZnQ0FA6depUbl/p6ek0btwY0H35FlG7dm0yMzP1x/369ePLL7/UC5GzZ8+W25+1tTVKpZKQkJBSuxT/zJ+vU177hIQEatWqxdixY5k1axYRERH6NgsXLsTS0pJp08qOeuXl5cXy5cvp3r077du3JyQkBGNjY+rUKb0DW6lUlmtdfFz69evHxo0bycrKAiA+Pp579+6VmnNx7Ozs+OOPP8jJyUEIwaFDh8r0Lzt79ixvvPEGu3fvxtr64Xfili1buHPnDjExMSxfvpxx48Y9llBLS0vD19eXJUuW0LVr13LrHThwgKVLl7J7925q1Xq2Y0NWZQLKhsDOQiWrALYKXbw20PnClVgClSTJFtgghPB5RNsaQajV3P3oY4zs7bFb/y1yc3NyMtI59N06rv1xlEYtW9H/7Q+wsm1ck8PkWuo1Pjj8AXGZcczsMJNxbcc9008TerRayLkPmQm6QLQZCZCZqHtlJD58n3O/dFuFCdS2gdq2YOcJLXrrEqebGMJQFLHm9+t8G3qT8V2aMbPvP0TYG6hSJEli586dTJ8+nc8++wwTExPs7e1ZtWoV3bp148SJE7i6uiJJEkuXLsXGxqZca46/vz/jx49n0aJF+Pr66st79erFkiVLcHNzY86cOcyfP5/p06fj4uKCVqvFwcGBPXv2lOpvzJgxvPzyyzg7O9OhQwdat25d4Vzq1atH165dadeuHQMGDGD27Nlltr9w4QKzZs1CJpOhVCpZt25diX5Wr17NpEmT8Pf31/uoFeHl5UVsbCzdu3dHLpfTtGnTcsc1depUXFxccHd3Z/HixRWO/VH07duXy5cv06VLF0AXuiQwMJAWLVqUmHPxDQadO3dm2LBhuLu7o1AoaN++PVOnTgVK+qzNmjWLrKwshg8fDuhE3u7duyscT3Gfs507d/Luu++SnJyMr68vbm5u/Pbbb6xZs4br16+zcOFC/bJrUFAQ1tbWJdq/88475Ofn412Y9tHT05Ovv/76b92vmkIqegJ5HujQoYMob1fJk6Dgzh1kpqYoGjTgxpmTBH3zJXlZWbw4fDQdB72CTF6z6YR239jNf078h9pGtVnWYxkeDctdZX66yM/SCbDMhJLCK6NQmGUm6n5q//wkKYG5tS7vZu1GYNFIJ8hq25R8b2r5j7WYVYaNR2+xcM8lXnFvwrJhLshkhnv1PHD58uVyd9MZMGCg+inrMylJ0hkhxCN3U1SlZe25w8jOjvycHILXreLi4WAa2Nkz7P/+Q4NmNeuInq/JZ8mpJfxy7Rc62nRkafel1Dd9CkItaNSQlfQnC1hxy1ihEMvPKN3W2KLQGtYImnX9kxCz1ZWbW4NcWf3zeo7YdjqWhXsuMaCdDZ+94mwQagYMGDDwFGIQa4/BnahzHFi3iqz79+k8ZARdho1CrqhZsRCXGceMIzO4dP8Sk9pN4t3276KQVfGvVQjITX1oDcu8W4YQS4Sse8CfLLcyRaElzAYatIIWvR4uUVo0enjOuHbVzsEAe88n8uGO83R3bMCqkW4o5I/lwmrAgAEDBqoJg1irJFqNhkPfrUOhVDJy4VJsHSv2cagOQuNCmRM2ByEEX/T6gl52vR7d6FGo8h4uO5Zalkx8eE5dxm4jU6uHVi8b54fv9UuUjaBWfZAZREFNE3LlHtN/OotHM0u+GeuBsaJml/ANGDBgwED5GMRaJZHJ5QyevQBzSyuUxjUbwV6j1fBV5Fesv7Ce1latWdFjBU0tmv71DtUFcOVXCk5txOjO0dKn5SbkmVhTYGpNQW0nVA17ozazQWNug8ascKmydkPkRrUwkstQyCUUcqnwvQyFTPfesMT2dPDHzfu8GXiGVja1+W5CR0yNDELNgAEDBp5mDGLtMbC0KZ1Oo7q5n3uf2WGzOZl4kqEthzKn0xxMFH9RPN6/ARGb4ewWyEnhgWTNL2o/bouG3BVW3BVWJIm6ZGAG2eUJLRVwp/BVMTIJFHLZQ0Enk2Ekl3SCTi/uisqLRF9hHdmf6hT1IytWp3g/MgmlQoaysJ1SLkMplzBSyGhUxxT7emb/SJFyLjaNyZvDsbOqxfeTOmNhYvD5M2DAgIGnHYNYe4aIvBfJjCMzSM9PZ+GLCxnScsjjd6IugKt7ITwAbh0BSY7WcQBLU7qw8a49G8Z3ZrxdXVQagVqjRaUVqNRa1FotKo1ApdE+PKcRqLRa1PryYu+1RXWK6heVF69frI62WJ/F+sku0KAu0W9Z7XXjeNyNzY3rmuJQ30z/at7AjOb1zWlsaYr8ObQCXr2byfiAU1iaKQmc3BkrM6OaHpIBAwYMGKgEBrH2DCCEIPByICvCV9DIvBGBPoG0tnpMn7kHN+HMZojcAtnJUMcOes9DuI1h5v577IiPZ8UIV7o7NqiaSVQDGu2jRWO+Wktcag63krO5maJ7/S8ynsw8tb4fI7kMu3q1aF7fDIcGZjSvb0bzBuY41DejnpnRMxm3LiYlm7HfncRYIWPL6540tHj+k9EbqHnu3r3L9OnTOX36NHXr1qVhw4asWrUKR0fHMuvHxMQwcOBAoqKiOHz4MMuXLy8zTloRkZGRJCQk4OPj88TGbG5urg8QW0RaWhpbt27l7bff/kt92tvbEx4eTv36f32X/uTJk/nggw9o27Ytn3zyCXPnzgVK3rMnyaPmvHLlSjZs2IAkSTg7OxMQEICJSem/K9u2beOjjz5CkiRcXV3ZunWr/lxGRgZt27Zl8ODB+vyuxZk1axa//vorRkZGtGjRgoCAAOrWrfvkJvkMYRBrTznZqmwWHFtA0O0gejXtxaJui7AwsqhcY40KruyFMwFw8zBIcnDsDx0m6gLDyuSsDLrKjrPxzPB2ZKj7s50HUi6TkMvkmCgrXt50a1rywy6E4EF2ATdTsvUi7lZKFjeTszl8NZkCjVZft7aJooR4K26Vq2X0dH6cEtNzGbPhJGqNlm1vdMGu3rMdydvAs4EQgiFDhjB+/Hh+/PFHAM6dO0dSUlK5Yu1xiYyMJDw8/LHEmlqtRqF4vM9qWloaa9eu/cti7UlQPOdlcbFWVVQ05/j4eL744gsuXbqEqakpI0aM4Mcff2TChAkl6kVHR/Ppp59y7NgxLC0tuXfvXonz8+fPL5Wovjje3t58+umnKBQKZs+ezaeffspnn332ROb3rPF0frsYAOB66nXeP/w+dzLv8L7H+0x0mlg5q86DWxDxPZwNhOx7YNFEl5y8/VjdDs1Ctp2O5Yvfr/Nqh6a80/uFKpzJ040kSdQzN6aeuTEd7a1KnNNoBQlpudxIzuJWSrb+derWA3aejS9R18bCRCfe9NY4Mxzqm9PU0rTGwmKkZOUzZsNJMnJVbJ3iScuGhpAo/0TufvIJ+ZevPNE+jdu0xqYCwRASEoJSqeTNN9/Ul7m6ugI6Iefv78/+/fuRJIl58+bx6quvltvXqVOneO+998jLy8PU1JSAgAAcHBxYsGABubm5HD16lDlz5jBw4EDeffddoqKiUKlUfPTRR/j5+bFp0yZ27NhBVlYWGo2GvXv34ufnR2pqKiqVikWLFuHn51fu9T/88ENu3LiBm5sb3t7e/Pvf/y6zfXZ2NiNGjCAuLg6NRsP8+fNLzCs3N5ehQ4cydOhQpkyZoi//+eefOXHiBCtWrGD16tWsXr2amzdvcvPmTV577TWOHTtGz549Wb58Ob/88gu5ubm4ubnh5OTE4sWL0Wg0TJkyhePHj9O4cWN27dqFqalpiTlMmDABCwsLwsPDuXv3LkuXLmXYsGEALFu2jG3btpGfn8+QIUP4+OOPS825eAYD0Ine3NxclEolOTk52NqW9ulev34906ZNw9LSEqBEyqkzZ86QlJRE//79KS+Yfd++ffXvPT09n1iu2WcRg1h7Stlzcw8LTyyklqIWG/puoKNNx4obaFRwdb/Oinbjd5BkOiuax0R4oQ/ISlqbjlxLZs7OC3i1rM+iIe2eyaW96kAuk2hqVYumVrXo+acsTHkqDTH3H1rjbibrLHL7LiSSlvMw24JCJj1cVi1mlWte34wGtY2r7N6n56oY990pEtJy+e/rnXFuYkivZaD6iIqKwsOj7CwqO3bsIDIyknPnzpGSkkLHjh0rtLC0bt2asLAwFAoFwcHBzJ07l+3bt7Nw4ULCw8P1S2hz586ld+/ebNy4kbS0NDp16sRLL70EQEREBOfPn8fKygq1Ws3OnTuxsLAgJSUFT09PBg0aVO5nccmSJURFRREZGQlQbvsDBw5ga2vL3r17AV0O0SKysrIYOXIk48aNY9y4cSX69/Ly0qefCgsLo169esTHxxMWFlbqvixZsoQ1a9boxxITE0N0dDQ//PAD69evZ8SIEWzfvr3MpOqJiYkcPXqUK1euMGjQIIYNG0ZQUBDR0dGcOnUKIQSDBg0iNDS01JyL07hxY2bOnImdnR2mpqb07dtXL6yKp5u6du0aoEtir9Fo+Oijj+jfvz9arZYZM2YQGBhIcHBwib6Lp4sqzsaNGysU9M87BrH2lFGgKWDp6aX8dPUn3K3dWd5jOQ1qVeBHlnq7cEdnoC5bgEUT6DlXZ0WrU3ae0ksJGbwdeAbHhrVZO8YdpSEY6l/CRCmntY0FrW1KL0unFi2rpmRzs5hVLiw6hXz1w2VVc2NFmZsc7OvXovbf2KmZU6Bm0qbTRN/LZMP4jqUshgb+WVRkAasJjh49yqhRo5DL5TRs2JAePXpw+vRpXFxcyqyfnp7O+PHjiY6ORpKkcpOYBwUFsXv3bpYvXw5AXl4ed+7odqp7e3tjZaX7HAghmDt3LqGhochkMuLj40lKSsLGxqZS4y+vvbOzMzNmzGD27NkMHDgQLy8vfRs/Pz/8/f0ZM2ZMqf5sbGzIysoiMzOT2NhYRo8eTWhoKGFhYQwdOvSR43FwcMDNzQ0ADw+PcnOsDh48GJlMRtu2bUlKStLfs6CgINq3bw/oRGV0dDR2dnblXi81NZVdu3Zx69Yt6taty/DhwwkMDGTs2LH6XJ2gE7XR0dEcPnyYuLg4unfvzoULFwgMDMTHx4cmTUq73hRf7i1i8eLFKBSKMu/dPwWDWHuKSMhKYMbhGUTdj2Ki00T+5f6vsrMRaFRw7QCc2QTXD+nyXrbsBx4ToKV3KStacRLTc5m06TQWpkoCJnT8W4LAQPlYmhnhYWaERzPLEuVarSAhPVcv3m4WWuXOxqby6/mEEjtaG9Q2LracqltSbd7AjKaWtTBSlC+w81Qapn5/hrN3Ulk7xp0ez/CmEQPPLk5OTk9s2Wr+/Pn06tWLnTt3EhMTQ8+ePcusJ4Rg+/bttGpV0gx+8uRJzMzM9MdbtmwhOTmZM2fOoFQqsbe3Jy8vr9LjaA74RgAAIABJREFUKa+9o6MjERER7Nu3j3n/3969h0VZrY0f/y5hfoqa59Q8JZqHLadBEI9DiqftIQleMk+pHbT2a+3aFSgklryZhGaZ7mqrKZpmVuqrpfGSRoFpWwERSUvSPOEJTAwUEXD9/hicDTKDouCg3p/r8nJmzbPWcz/zhN2s9ay1pk+nf//+zJgxAzD3LsXExDBmzBirPXi9evVi2bJldOrUCZPJxNKlS9mxYwfvvPPOdeOpWbOm5bWDgwN5eVYWLb/muKv7gmutCQ0N5dlnny11rK2ED2DLli04Oztz//3mf1sCAwPZvn17md68Vq1a0b17dwwGA87OznTs2JH09HR27NhBQkICH3zwAbm5uVy+fJm6desSGRlZ5lzR0dF8/fXXbN269Z4eAZJkrZpIOJ5A6LZQiq4U8V7f9+j/YP+yB2UfNc/o3L0Sck+Zt2h6eCp0fQLqX39ywJ+XCnhy2S4u5Bfyxd960ry+zAi83WrUULRqWJtWDWtj6lA6ibpUUMTRPy4WD6f+Z5JD7M+nOXvhsuU4hxqK1g2dygypOt9fhyZ1a/LC6t1s+y2Ldx7z4K+uD9zuSxQCAD8/P8LCwli0aBGTJ08GIDU1lfPnz2MymfjXv/7FhAkT+OOPP4iPj2fOnDk2E6bz58/TsqV5pCA6OtpSft9995GTk2N5P3jwYBYsWMCCBQtQSrF7925Lj9G17TVt2hSDwUBcXBxHjhwp91quPY+t+idOnKBRo0aMGzeOBg0alOolioiIICIigilTpvDBBx+UOYfJZGLGjBnMmDEDT09P4uLicHJyon79so8vGAwGCgoKMBhu/ZftwYMHEx4eztixY6lbty4ZGRkYDIYy11xSmzZt+Omnn7h48SJOTk5s3bq1zLAlmHvyVq9ezZNPPklWVhYHDhygXbt2rFq1ynJMdHQ0iYmJVhO1mJgYoqKi+OGHH6hd+96eGCXJmp0VXSniwz0fsih1ER0admBe33k8WO/BEgcUluhF22LuRXtoIHi/Z/7b4cZuYUHRFf57ZTK/nckl+kkfq0N3wr5qGRzo2Ow+OlqZBHD+YgG/n/3PkOrVmas/HfqDvIIiy3EGB0VBkSbC34X/8rqzZ/eKO5tSivXr1/PSSy/x9ttvU6tWLdq2bct7771Hnz592LFjBx4eHiiliIqKonnz5jZ7c0JCQpgwYQJvvvkmw4YNs5T369ePyMhIjEYjoaGhhIeH89JLL+Hu7s6VK1dwdna2uvTH2LFjeeSRR3Bzc8Pb25vOnctfCqlx48b07t0bV1dXhgwZwtSpU63W37t3L8HBwdSoUQODwcCHH35Yqp358+fz1FNPERISYnlG7SqTycSxY8fw9fXFwcGB1q1b24xr8uTJuLu707VrV2bNmlVu7NczaNAg9u/fT8+ePQHz0iUrV66kffv2pa655ASD7t27ExQURNeuXXF0dMTT09OSkJd8Zm3w4MHExsbSpUsXHBwcmDNnDo0bNy43npLPrD3//PPk5+czcOBAwDzJ4KOPPrql671TKV3RlUSrMW9vb21rVkl1dO7SOabGT2XHyR34t/fntR6v4eRYPIMn+1jxjM5PzPtx3vcAdB0Pnk9Ag4ptLaW1JvjLVL5MOs6cIHce876FralEtXLliuZ0zqVSkxzcWtUjwFMStXvd/v37+ctf/mLvMIQQxaz9TCqlkrTWZbslryE9a3ayJ3MPr3z/CucuneONnm8Q2CEQdaUIftlsntGZ/q35wA4DYdg75mfSbrAX7Vrzt6bzZdJxXhrQQRK1u0yNGooH6jvxQH0nej108wtuCiGEqL4kWbvNtNZ8+sunzE2cS7Pazfhk6Cd0cawH38+G5E8g5wTUbQ6+r5p70hrYnpFzI75IPMZ7W9IJ8mrFi/07VNJVCCGEEOJ2kWTtNrpYcJE3tr/BN4e/oW+rh3mzeT/q/99MSI8FreGhATB0jnl9tJvsRStpW3oWoev20uehJrwV4HZPz6QRQggh7lRVmqwppQ4DOUARUKi19lZKvQFMAjKLDwvTWm+2UvevwHzAAViitS47VeQOcjD7IP/4/h8cOX+EFxt48NSe76iR8AnUbQZ9Xjb3ojV88PoN3aD9J//kuZVJPNS0Lh+M61ruUg9CCCGEqL5uR89aP6111jVl72qt59qqoJRyAP4JDASOA7uUUhu11vuqMM4q883BTby+fQZOV4pYfOo0Pr8fMe/NOeTt4l60yl3r7NT5Szy5bBd1ajqw7Mlu1JO11IQQQog7VnUdBvUBftNaHwJQSn0G+AN3VLJWkH2EOd+9wuqcX/G8dIk5udDM+7+Le9HaVsk5cy4V8GT0LnLzC/n82Z48UN/p+pWEEEIIUW1V9diYBmKVUklKqcklyp9XSqUqpZYqpRpaqdcSOFbi/fHisjKUUpOVUolKqcTMzExrh9xeV4rgQCynPv0vJn4+kNU5vzKe+nzc+22avbQP+s+oskStoOgK/70qmQOnc/hgbFe6tJC11IQQ9nPq1ClGjRpF+/bt8fLyYujQoZb9Iq05fPgwrq6uAHz//fcMHz683PZTUlLYvLnMUzS3pG7dumXKsrOzrS5ke6Patm1LVta1A0wV88wzz7Bvn7m/4q233rKUl/zOKtP1rvndd9/FxcUFV1dXRo8ebXVB448++gg3NzeMRiN9+vSxxP/tt9/i5eWFm5sbXl5efPfdd1bPERwcTOfOnXF3dycgIIDs7OzKubg7UFUna3201l2BIcAUpZQv8CHQHjACJ4Hr76VRDq31Iq21t9ba++rWF3bx50n4YQ7M92D7+id4LP9XDtaqwzyvaQRP2IbBNbDShztL0lozfX0aCelZzA5ww1e2GBJC2JHWmoCAAPr27cvBgwdJSkpi9uzZlj0pK8PNJGuFhYUVPs+tJmuVYcmSJXTp0gUonaxVlfKuOSMjg/fff5/ExETS0tIoKiris88+K3PcmDFj2Lt3LykpKYSEhPDyyy8D0KRJE7766iv27t3L8uXLeeKJJ6yeZ+DAgaSlpZGamkrHjh2ZPXt25V3gHaZKh0G11hnFf59RSq0HfLTW8Vc/V0otBsouLw0ZQMkFwVoVl1UvV4rgYJx5XbRfv+GKLuJfzh58SA3a12/Hu/3eo239trcllIXf/caaxGP83e8hRnaTtdSEEP+R8PkBso7lVmqbTVrXxTSyo83P4+LiMBgMPPfcc5YyDw8PwJzIhYSE8M0336CUYvr06Tz++OM229q5cycvvvgily5dwsnJiWXLluHs7MyMGTPIy8tj27ZthIaGMnz4cF544QXS0tIoKCjgjTfewN/fn+joaNatW0dubi5FRUVs2rQJf39/zp07R0FBAW+++Sb+/v42zz9t2jQOHjyI0Whk4MCBvP7661brX7hwgZEjR3L8+HGKiooIDw8vdV15eXkEBgYSGBjIpEmTLOVffPEFO3bsYN68ecyfP5/58+dz6NAhDh06xBNPPMGPP/5I3759mTt3Ll9++SV5eXkYjUZcXFyYNWsWRUVFTJo0ie3bt9OyZUs2bNiAk1PpR2AmTpxIvXr1SExM5NSpU0RFRREUFATAnDlz+Pzzz8nPzycgIICZM2eWueaSOxiAOenNy8vDYDBw8eJFWrRoUeZ7q1fvP6M7Fy5csKxIUHILMBcXF/Ly8sjPzy+1dymYd1e4qkePHpW21+ydqMqSNaVUHaCG1jqn+PUgIEIp9YDW+mTxYQFAmpXqu4AOSilnzEnaKGBMVcVaYTmnzDsLJK2A80ehdhPOdZ9M6JUT/Ji5m0faPcL0HtOpbbg9e5mtSz7OO98eINCzJf8YaPsfTyGEuF3S0tLw8vKy+tm6detISUlhz549ZGVl0a1bN3x9fW221blzZxISEnB0dGTLli2EhYWxdu1aIiIiSExMZOHChQCEhYXh5+fH0qVLyc7OxsfHhwEDBgCQnJxMamoqjRo1orCwkPXr11OvXj2ysrLo0aMHI0aMsLm8UWRkJGlpaaSkpADYrB8TE0OLFi3YtGkTYN5D9Krc3FxGjRrF+PHjGT9+fKn2TSaTZfuphIQEGjduTEZGBgkJCWW+l8jISBYuXGiJ5fDhw6Snp7N69WoWL17MyJEjWbt2bZlN1QFOnjzJtm3b+OWXXxgxYgRBQUHExsaSnp7Ozp070VozYsQI4uPjy1xzSS1btuTVV1+lTZs2ODk5MWjQIEtiVXK7KYB//vOfzJs3j8uXL1sd7ly7di1du3a1JGolt5sqaenSpeUm9He7quxZawasL/6P3xH4VGsdo5T6RCllxPw822HgWQClVAvMS3QM1VoXKqWeB/4P89IdS7XWP1dhrNenNRzcConmXjR0ETg/DANnsrdJW17ZNo2svCxm9JxBUIeg27am2fbfspi6NpWe7RoT+V/uspaaEKKM8nrA7GHbtm2MHj0aBwcHmjVrxsMPP8yuXbtwd3e3evz58+eZMGEC6enpKKUoKCiwelxsbCwbN25k7lzzYgOXLl3i6NGjgHlIrVGjRoC5Zy8sLIz4+Hhq1KhBRkYGp0+fpnnz5jcUv636bm5uvPLKK0ydOpXhw4djMpksdfz9/QkJCWHs2LFl2mvevDm5ubnk5ORw7NgxxowZQ3x8PAkJCQQGBl43HmdnZ4xGIwBeXl4291h99NFHqVGjBl26dLEMR8fGxhIbG2vp7crNzSU9PZ02bWwvyH7u3Dk2bNjA77//ToMGDXjsscdYuXIl48aNIyIiotSxU6ZMYcqUKXz66ae8+eabLF++3PLZzz//zNSpU4mNjbWULVmypMz5Zs2ahaOjo9Xv7l5RZcla8UxODyvlVgentdYngKEl3m8GKvfJ0Vv17RvmHQZ6TgGviehG7Vjz6xre/vZpmjo15ZMhn+DSxOW2hfPrqRyeXZmEc5M6fPSEl6ylJoSoNlxcXCpt2Co8PJx+/fqxfv16Dh8+TN++fa0ep7Vm7dq1dOrUqVT5v//9b+rUqWN5v2rVKjIzM0lKSsJgMNC2bVurD8jbYqt+x44dSU5OZvPmzUyfPp3+/fszY8YMAHr37k1MTAxjxoyx+kt1r169WLZsGZ06dcJkMrF06VJ27NjBO+9c/7HuksOHDg4O5OXlXfe4q/uCa60JDQ3l2WefLXWsrYQPYMuWLTg7O3P1OfHAwEC2b99utTfvqlGjRvG3v/3N8v748eMEBASwYsUK2rdvb7NedHQ0X3/9NVu3br2nOyPk/+43Sil4fAW8vB8G/Q8X6z1A6LZQZv17Fj0f6Mnnj3x+WxO1039e4sllO3EyOLDsSR/qO8laakKI6sPPz4/8/HwWLVpkKUtNTSUhIQGTycSaNWsoKioiMzOT+Ph4fHx8bLZ1/vx5WrY0LwgQHR1tKb/vvvvIycmxvB88eDALFiywJCK7d++22V7Tpk0xGAzExcVx5MiRcq/l2vPYqn/ixAlq167NuHHjCA4OJjk52VInIiKChg0bMmXKFKvnMJlMzJ07F19fXzw9PYmLi6NmzZrUr1+/zLEGg8Fm72JFDR48mKVLl5Kba36mMSMjgzNnzpS55pLatGnDTz/9xMWLF9Fas3Xr1jIblAOkp6dbXm/atIkOHcxbHmZnZzNs2DAiIyPp3bu3zdhiYmKIiopi48aN1K59ex4rqq4kWauIRu3AsSaHzh9izKYxbD60mRc8X2Bh/4XUr1n2B6qq5OYX8lT0LrLzClg6sRstG8haakKI6kUpxfr169myZQvt27fHxcWF0NBQmjdvTkBAAO7u7nh4eODn50dUVFS5Q5AhISGEhobi6elZajZnv3792LdvH0ajkTVr1hAeHk5BQQHu7u64uLgQHh5utb2xY8eSmJiIm5sbK1asoHPnzuVeS+PGjenduzeurq4EBwfbrL937158fHwwGo3MnDmT6dOnl2pn/vz55OXlERISUuYcJpOJY8eO4evri4ODA61bt6ZPnz5W45k8eTLu7u6VMiw4aNAgxowZQ8+ePXFzcyMoKIicnJwy11xS9+7dCQoKomvXrri5uXHlyhUmTzavzjVjxgw2btwIwMKFC3FxccFoNDJv3jzLEOjChQv57bffiIiIwGg0YjQaOXPmDGB+Zi0xMRGA559/npycHAYOHIjRaCw1WeVeo67+BnI38Pb21ldvclWJORzD6z++Ti3HWkSaIunZomeVnu9ahUVXeHp5Itt+y2LJBG/6dWp6W88vhLgz7N+/32pvhxDCPqz9TCqlkrTW3jaqWFTXHQyqncIrhbyT+A4r96/E434P5j48l+Z1buxh1MqitSZ8Qxo/HMhkdqCbJGpCCCHEPUCStRvkoBw4c/EM4/4yjpe9XsZQhQvc2vLB9wdZvfMYU/q1Z7SP7Zk6QgghhLh7SLJ2g5RSRPlG4VDDwS7n/9/dGcz5v1/xN7bg1UGdrl9BCCGEEHcFmWBQAfZK1HYcPEvwl3vo0a4RUUGylpoQQghxL5FkrZpLP53Ds58k8mDjOvxrnDc1He2TMAohhBDCPiRZq8bO/HmJict28f8cHVg2sRv1a8taakIIIcS9RpK1aupCfiFPLd/FHxcus2xiN1o3urcXBBRC3HlOnTrFqFGjaN++PV5eXgwdOpQDBw7YPP7w4cO4uroC8P333zN8+PBy209JSWHz5srd6KZu3bplyrKzs/nggw9uus22bduSlZV1K2HxzDPPsG/fPgDeeustS3nJ76wyXe+a3333XVxcXHB1dWX06NFWd4CIj4+na9euODo6ltnN4ujRowwaNIi//OUvdOnSxeqOCfPmzaNLly64u7vTv3//6y5efDeTZK0aKiy6wgurd7PvxJ/8c6wnbq1u34K7QghRGbTWBAQE0LdvXw4ePEhSUhKzZ8+27ElZGW4mWSu5qO6NutVkrTIsWbKELl26AKWTtapS3jVnZGTw/vvvk5iYSFpaGkVFRXz22WdljmvTpg3R0dGMGTOmzGfjx48nODiY/fv3s3PnTpo2LbsUlaenJ4mJiaSmphIUFGR1MeF7hcwGrWa01ry+8We+++UMbz7qil/nZvYOSQhxh4uLXsSZI4cqtc2mD7aj38TJts8ZF4fBYCi16ryHh3m7aK01ISEhfPPNNyilmD59Oo8//rjNtnbu3MmLL77IpUuXcHJyYtmyZTg7OzNjxgzy8vLYtm0boaGhDB8+nBdeeIG0tDQKCgp444038Pf3Jzo6mnXr1pGbm0tRURGbNm3C39+fc+fOUVBQwJtvvom/v7/N80+bNo2DBw9iNBoZOHAgr7/+utX6Fy5cYOTIkRw/fpyioiLCw8NLXVdeXh6BgYEEBgYyadIkS/kXX3zBjh07mDdvHvPnz2f+/PkcOnSIQ4cO8cQTT/Djjz/St29f5s6dy5dffkleXh5GoxEXFxdmzZpFUVERkyZNYvv27bRs2ZINGzbg5FR6Z5uJEydSr149EhMTOXXqFFFRUQQFBQEwZ84cPv/8c/Lz8wkICGDmzJllrnnOnDml2issLCQvLw+DwcDFixdp0aJFme+tbdu2ANSoUbpfaN++fRQWFjJw4EDAem8mmHeouKpHjx6sXLnS5j2620myVs189MMhVv37KM893J5xPR60dzhCCHFT0tLS8PLysvrZunXrSElJYc+ePWRlZdGtWzd8fX1tttW5c2cSEhJwdHRky5YthIWFsXbtWiIiIkhMTGThwoUAhIWF4efnx9KlS8nOzsbHx4cBAwYAkJycTGpqKo0aNaKwsJD169dTr149srKy6NGjByNGjLA50z4yMpK0tDRSUlIAbNaPiYmhRYsWbNq0CTDvIXpVbm4uo0aNYvz48YwfP75U+yaTiaioKAASEhJo3LgxGRkZJCQklPleIiMjWbhwoSWWw4cPk56ezurVq1m8eDEjR45k7dq1VjdVP3nyJNu2beOXX35hxIgRBAUFERsbS3p6Ojt37kRrzYgRI4iPjy9zzSW1bNmSV199lTZt2uDk5MSgQYMYNGgQYN5uytvbmxEjRti8nwcOHKBBgwYEBgby+++/M2DAACIjI3FwcLBZ/+OPP2bIkCE227zbSbJWjWzcc4K3Y37hEY8WhAyWtdSEEJWjvB4we9i2bRujR4/GwcGBZs2a8fDDD7Nr1y7c3d2tHn/+/HkmTJhAeno6Simbm5jHxsayceNG5s6dC8ClS5c4evQoAAMHDqRRo0aAuWcvLCyM+Ph4atSoQUZGBqdPny53f9KSbNV3c3PjlVdeYerUqQwfPhyTyWSp4+/vT0hIiNX9PJs3b05ubi45OTkcO3aMMWPGEB8fT0JCAoGBgdeNx9nZGaPRCICXl5fV578AHn30UWrUqEGXLl0sw9GxsbHExsbi6ekJmJPK9PR02rSxvfD6uXPn2LBhA7///jsNGjTgscceY+XKlYwbN46IiIjrxltYWEhCQgK7d++mTZs2PP7440RHR/P0009brb9y5UoSExP54Ycfrtv23UqeWasm/n3oLK9+vgefto2Y+5g7NWrIWmpCiDuXi4sLSUlJldJWeHg4/fr1Iy0tja+++srqw+xgTqLWrl1LSkoKKSkpHD161LIXY506dSzHrVq1iszMTJKSkkhJSaFZs2Y227TGVv2OHTuSnJyMm5sb06dPL5V49O7dm5iYGGztx92rVy+WLVtGp06dMJlMJCQksGPHDnr37n3deGrWrGl57eDgYPO5vJLHXY1Da01oaKjlO/vtt994+umnyz3fli1bcHZ25v7778dgMBAYGMj27duvG+dVrVq1wmg00q5dOxwdHXn00UdJTk62ea5Zs2axcePGUvHfayRZqwZ+O5PL5E+SaNXIiUXjvWQtNSHEHc/Pz4/8/HwWLVpkKUtNTSUhIQGTycSaNWsoKioiMzOT+Ph4fHx8bLZ1/vx5WrZsCUB0dLSl/L777iMnJ8fyfvDgwSxYsMCSiOzevdtme02bNsVgMBAXF3fdWYbXnsdW/RMnTlC7dm3GjRtHcHBwqQQkIiKChg0bMmXKFKvnMJlMzJ07F19fXzw9PYmLi6NmzZrUr192gpnBYLDZu1hRgwcPZunSpeTm5gLmyQNnzpwpc80ltWnThp9++omLFy+itWbr1q1lNigvT7du3cjOziYzMxOA7777zjJ5oqTdu3fz7LPPsnHjRqsTEO4lkqzZWWZOPhOX7cTgoFj+pA8Nav8/e4ckhBC3TCnF+vXr2bJlC+3bt8fFxYXQ0FCaN29OQEAA7u7ueHh44OfnR1RUVLlDkCEhIYSGhuLp6Vmq16hfv37s27cPo9HImjVrCA8Pp6CgAHd3d1xcXAgPD7fa3tixY0lMTMTNzY0VK1bQuXPncq+lcePG9O7dG1dXV4KDg23W37t3Lz4+PhiNRmbOnMn06dNLtTN//nzy8vKszmo0mUwcO3YMX19fHBwcaN26NX369LEaz+TJk3F3d7c6pFpRgwYNYsyYMfTs2RM3NzeCgoLIyckpc80lde/enaCgILp27YqbmxtXrlxh8mTzUPuMGTPYuHEjALt27aJVq1Z88cUXPPvss7i4uADm3r+5c+fSv39/3Nzc0FpbJlyUrB8cHExubi6PPfYYRqOx3Ofg7nbKVpdspTSu1GEgBygCCrXW3kqpOcAjwGXgIPCk1jr7Rupe73ze3t46MTGx8i6gil28XMioRT+RfjqXzyb3wKN1A3uHJIS4S+zfv79CvR1CiKpl7WdSKZV0I/nN7ehZ66e1NpYI5lvAVWvtDhwAQitQ965RdEXz99W7Scs4z/ujPSVRE0IIIYRVt30YVGsdq7W+2o/9E9Dqdsdgb1pr3tj4M1v2n+GNES4M7CJrqQkhhBDCuqpO1jQQq5RKUkpZmzv+FPDNTda9Yy1OOMQnPx1hsm87xvdsa+9whBBCCFGNVfU6a3201hlKqabAt0qpX7TW8QBKqdeAQmBVReuWVJzITQbKXRemuvg69QRvbf6FYW4PMO2v5T/UKoQQQghRpT1rWuuM4r/PAOsBHwCl1ERgODBW25jhYKuuleMWaa29tdbe999/f6VfQ2XadfgPXv58D94PNuSdkR6ylpoQQgghrqvKkjWlVB2l1H1XXwODgDSl1F+BEGCE1vpiRepWVay3w8HMXCatSKRlAycWj/emlkHWUhNCCCHE9VVlz1ozYJtSag+wE9iktY4BFgL3YR7aTFFKfQSglGqhlNp8nbp3pKxc81pqDkoR/WQ3GtaRtdSEEHe/U6dOMWrUKNq3b4+XlxdDhw7lwIEDNo8/fPgwrq6uAHz//fcMHz683PZTUlLYvHlzucdUlLVNxbOzs/nggw9uus22bduSlZV1K2HxzDPPsG/fPgDeeustS3nJ7+x6VqxYgaurK25ubnh6elq25Zo4cSItW7YkPz8fgKysLMsm7IcPH0YpxYIFCyztPP/886UWJ74qPz+fxx9/nIceeoju3bvb3PZKVFyVJWta60Naa4/iPy5a61nF5Q9prVsXL8lh1Fo/V1x+Qms9tLy6d6K8y0U8vTyRzJx8lkzw5sHGda5fSQgh7nBaawICAujbty8HDx4kKSmJ2bNnW/akrAw3k6zZ2oqpPLearFWGJUuWWFb5L5ms3ahvvvmG9957j9jYWPbu3ctPP/1UancEBwcHli5darVu06ZNmT9/PpcvXy73HB9//DENGzbkt99+4x//+AdTp06tcJzCOtnBoAoVXdH8/bPdpB7PZv4oTzzbNLR3SEKIe1D2Vwc586/USv2T/dXBcs8ZFxeHwWDgueees5R5eHhgMpnQWhMcHGzp5VmzZk25be3cuZOePXvi6elJr169+PXXX7l8+TIzZsxgzZo1lh0MLly4wFNPPYWPjw+enp5s2LABMG9RNWLECPz8/Ojfvz+5ubn079/fsgL/1eNsmTZtGgcPHsRoNFpW1bdW/8KFCwwbNgwPDw9cXV3LXFdeXh5Dhgxh8eLFpcq/+OILXn75ZcC8y0G7du0AOHTokGVv0L59+5KYmMi0adPIy8tWfDLPAAAMJUlEQVTDaDRadjAoKipi0qRJuLi4MGjQIPLy8spcw+zZs5k7dy4tWrQAzPuEXt01AOCll17i3XfftZrM3n///fTv35/ly5eX+z1t2LCBCRMmABAUFMTWrVtt7oUqKqaqZ4Pes7TW/M/X+/h232neeKQLg11sb6UihBB3m7S0NLy8vKx+tm7dOlJSUtizZw9ZWVl069YNX19fm2117tyZhIQEHB0d2bJlC2FhYaxdu5aIiAgSExNZuHAhAGFhYfj5+bF06VKys7Px8fFhwIABACQnJ5OamkqjRo0oLCxk/fr11KtXj6ysLHr06MGIESNQyvqkr8jISNLS0khJSQGwWT8mJoYWLVqwadMmwLyH6FW5ubmMGjWK8ePHM378+FLtm0wmoqKiAEhISKBx48ZkZGSQkJBQ5nuJjIxk4cKFllgOHz5Meno6q1evZvHixYwcOZK1a9cybty4G74fYF5NoU+fPnzyySc88sgjZT6fOnUqQ4YM4amnnrLZRkZGBq1btwbA0dGR+vXrc/bsWZo0aWKzjrgxkqxVkY+3/U709sM83ceZib2d7R2OEOIe1uCR9vYOoZRt27YxevRoHBwcaNasGQ8//DC7du3C3d3d6vHnz59nwoQJpKeno5SyuYl5bGwsGzdutDyLdenSJY4ePQrAwIEDadSoEWD+ZTosLIz4+Hhq1KhBRkYGp0+fLnd/0pJs1Xdzc+OVV15h6tSpDB8+HJPJZKnj7+9PSEiI1f08mzdvTm5uLjk5ORw7dowxY8YQHx9PQkICgYGB143H2dkZo9EIgJeX100/KxYaGoq/vz/Dhg0r81m7du3o3r07n3766U21LW6NDINWgW/2nmTW5v0McW3Oa0Nlbz4hxL3HxcWFpKSkSmkrPDycfv36kZaWxldffcWlS5esHqe1Zu3ataSkpJCSksLRo0ctezHWqfOf54VXrVpFZmYmSUlJpKSk0KxZM5ttWmOrfseOHUlOTsbNzY3p06cTERFhqdO7d29iYmJsDgv26tWLZcuW0alTJ0wmEwkJCezYscMyDFqemjVrWl47ODhYHcq8kfvRoUMHjEYjn3/+udXPw8LCePvtt21eQ8uWLTl27Bhg7n08f/48jRs3vm784vokWatkSUf+4KU1KXi2bsC7jxtlLTUhxD3Jz8+P/Px8Fi1aZClLTU0lISEBk8nEmjVrKCoqIjMzk/j4eHx8rC6lCZh71lq2bAlQahbifffdR05OjuX94MGDWbBggSWZ2L17t832mjZtisFgIC4ujiNHjpR7Ldeex1b9EydOULt2bcaNG0dwcDDJycmWOhERETRs2JApU6ZYPYfJZGLu3Ln4+vri6elJXFwcNWvWLDUJ4CqDwWCzd9GW0NBQgoODOXXqFACXL19myZIlZY577bXXLD2T1+rcuTNdunThq6++svr5iBEjLM+1ffnll/j5+dkcWhYVI8laJfo96wLPLE/kgfq1ZC01IcQ9TSnF+vXr2bJlC+3bt8fFxYXQ0FCaN29OQEAA7u7ueHh44OfnR1RUVLlDkCEhIYSGhuLp6Vmq16hfv37s27fPMsEgPDycgoIC3N3dcXFxITw83Gp7Y8eOJTExETc3N1asWEHnzuXvJtO4cWN69+6Nq6srwcHBNuvv3bsXHx8fjEYjM2fOZPr06aXamT9/Pnl5eYSEhJQ5h8lk4tixY/j6+uLg4EDr1q3p06eP1XgmT56Mu7u71SFVW4YOHcrzzz/PgAEDcHFxoWvXrvz5559ljrv6mS2vvfYax48ft/rZ008/zdmzZ3nooYeYN28ekZGRNxyfKJ+6m2ZqeHt768TERLuc+2xuPoEfbifnUiHr/taLtk1kiQ4hhP3s37/fMgQohLA/az+TSqkkrbX39epKz1oluFRQxDMrEjl1/hKLx3tLoiaEEEKISiOzQW9R0RXNi5/tJuVYNh+M6YrXg7KWmhBCCCEqj/Ss3aJZm/bzfz+fZvqwLgxxe8De4QghhMXd9JiLEHeyW/1ZlGTtFizd9jtLf/ydib3a8nQfWUtNCFF91KpVi7Nnz0rCJoSdaa05e/YstWrVuuk2ZBj0JsWkneJ/Nu1jUJdmhA/vYu9whBCilFatWnH8+HEyMzPtHYoQ97xatWrRqlWrm64vydpNSD56jhc/241HqwbMH+WJg6ylJoSoZgwGA87O0uMvxN1AhkEr6MhZ81pqzerVYskEb5z+n6ylJoQQQoiqI8laBfxx4TITl+3iitZEP9mNJnVrXr+SEEIIIcQtkGHQG1RYdIVJKxLJyM7j02e60+7+uvYOSQghhBD3gLtqBwOlVCZQ/iZv4nqaAFn2DkLcErmHdz65h3c2uX93vtt1Dx/UWt9/vYPuqmRN3DqlVOKNbH0hqi+5h3c+uYd3Nrl/d77qdg/lmTUhhBBCiGpMkjUhhBBCiGpMkjVxrUX2DkDcMrmHdz65h3c2uX93vmp1D+WZNSGEEEKIakx61oQQQgghqjFJ1oQQQgghqjFJ1gQASqnWSqk4pdQ+pdTPSqkX7R2TqDillINSardS6mt7xyIqTinVQCn1pVLqF6XUfqVUT3vHJCpGKfWP4n9D05RSq5VStewdkyifUmqpUuqMUiqtRFkjpdS3Sqn04r8b2jNGSdbEVYXAK1rrLkAPYIpSqoudYxIV9yKw395BiJs2H4jRWncGPJB7eUdRSrUE/g54a61dAQdglH2jEjcgGvjrNWXTgK1a6w7A1uL3diPJmgBAa31Sa51c/DoH8/8kWto3KlERSqlWwDBgib1jERWnlKoP+AIfA2itL2uts+0blbgJjoCTUsoRqA2csHM84jq01vHAH9cU+wPLi18vBx69rUFdQ5I1UYZSqi3gCfzbvpGICnoPCAGu2DsQcVOcgUxgWfFQ9hKlVB17ByVunNY6A5gLHAVOAue11rH2jUrcpGZa65PFr08BzewZjCRrohSlVF1gLfCS1vpPe8cjboxSajhwRmudZO9YxE1zBLoCH2qtPYEL2HnoRVRM8XNN/pgT7xZAHaXUOPtGJW6VNq9xZtd1ziRZExZKKQPmRG2V1nqdveMRFdIbGKGUOgx8BvgppVbaNyRRQceB41rrqz3aX2JO3sSdYwDwu9Y6U2tdAKwDetk5JnFzTiulHgAo/vuMPYORZE0AoJRSmJ+V2a+1nmfveETFaK1DtdattNZtMT/Q/J3WWn6jv4NorU8Bx5RSnYqL+gP77BiSqLijQA+lVO3if1P7I5NE7lQbgQnFrycAG+wYiyRrwqI38ATmHpmU4j9D7R2UEPeYF4BVSqlUwAi8Zed4RAUU94p+CSQDezH/P7ZabVskylJKrQZ2AJ2UUseVUk8DkcBApVQ65h7TSLvGKNtNCSGEEEJUX9KzJoQQQghRjUmyJoQQQghRjUmyJoQQQghRjUmyJoQQQghRjUmyJoQQQghRjUmyJoS4qymlikosR5OilKq0XQGUUm2VUmmV1Z4QQljjaO8AhBCiiuVprY32DkIIIW6W9KwJIe5JSqnDSqkopdRepdROpdRDxeVtlVLfKaVSlVJblVJtisubKaXWK6X2FP+5uo2Qg1JqsVLqZ6VUrFLKqfj4vyul9hW385mdLlMIcReQZE0IcbdzumYY9PESn53XWrsBC4H3issWAMu11u7AKuD94vL3gR+01h6Y9+z8ubi8A/BPrbULkA38V3H5NMCzuJ3nqurihBB3P9nBQAhxV1NK5Wqt61opPwz4aa0PKaUMwCmtdWOlVBbwgNa6oLj8pNa6iVIqE2iltc4v0UZb4FutdYfi91MBg9b6TaVUDJAL/C/wv1rr3Cq+VCHEXUp61oQQ9zJt43VF5Jd4XcR/ngUeBvwTcy/cLqWUPCMshLgpkqwJIe5lj5f4e0fx6+3AqOLXY4GE4tdbgb8BKKUclFL1bTWqlKoBtNZaxwFTgfpAmd49IYS4EfKbnhDibueklEop8T5Ga311+Y6GSqlUzL1jo4vLXgCWKaWCgUzgyeLyF4FFSqmnMfeg/Q04aeOcDsDK4oROAe9rrbMr7YqEEPcUeWZNCHFPKn5mzVtrnWXvWIQQojwyDCqEEEIIUY1Jz5oQQgghRDUmPWtCCCGEENWYJGtCCCGEENWYJGtCCCGEENWYJGtCCCGEENWYJGtCCCGEENXY/wc1DrVO+pow8gAAAABJRU5ErkJggg==\n",
      "text/plain": [
       "<Figure size 720x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "architectures = [\n",
    "    (\n",
    "        nn.Conv2d(1, 20, 5, 1),\n",
    "        nn.Conv2d(20, 50, 5, 1),\n",
    "        nn.Linear(4*4*50, 500),\n",
    "        nn.Linear(500, N_FONTS)\n",
    "    )\n",
    "]\n",
    "\"\"\"\n",
    ",\n",
    "    (\n",
    "        nn.Conv2d(1, 30, 4, 1),\n",
    "        nn.Conv2d(30, 100, 4),\n",
    "        nn.Linear(100 * 4 * 4, 1000),\n",
    "        nn.Linear(1000, 100),\n",
    "        nn.Linear(100, N_FONTS)\n",
    "    )\n",
    "\"\"\"\n",
    "for i, architecture in enumerate(architectures):\n",
    "    model = CNNet(architecture, quadratic_model)\n",
    "    _, test_perfs_font = evaluate_resistance(model)\n",
    "    \n",
    "    resistance_reports[f\"Collateral task with CNN {i}\"] = test_perfs_font\n",
    "    \n",
    "show_results(resistance_reports, title=\"Resistance of CNNs with CNN protection\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAmEAAAEWCAYAAAAuOkCvAAAABHNCSVQICAgIfAhkiAAAAAlwSFlzAAALEgAACxIB0t1+/AAAADl0RVh0U29mdHdhcmUAbWF0cGxvdGxpYiB2ZXJzaW9uIDMuMC4zLCBodHRwOi8vbWF0cGxvdGxpYi5vcmcvnQurowAAIABJREFUeJzs3Xd8VFXex/HPjxRqIBBCTwi9l0Doggioi4qurl0UwbqPq7LugwXL7vqsa1ldRV3dRRREQUGRVVfFCorSO0hNMCShJtQkkH6eP2ZgI6YMkGFSvu/Xi1cyd+7c+c1kcvPlnHPPMeccIiIiInJ2VQt0ASIiIiJVkUKYiIiISAAohImIiIgEgEKYiIiISAAohImIiIgEgEKYiIiISAAohIkEgJn9aGZDA12HP5jZb81sr5llmFlEoOspK6X9zMxsgZndehZLqvS8n6HWga5DxF8UwkRKYGaJZnbM+8dgj5lNM7M6Z3pc51wX59yCUp47xsycmQWf6fOdLWYWAvwduMA5V8c5t7+IfULN7E9mts3MMr3v8RtmFuO9f4GZZZlZVKHHjDCzxEK3E81sn5nVLrTtVjNb4K/XVvhn5q3/7TM5npnVNbMXzCzJ+/lK8N5u6L2/1Nfo/XysN7Nqhbb9xcymnUltp/Aabjaz78voWL8Isd7P0PayOL5IeaQQJlK6Uc65OkBPIBZ4KMD1lGeNgRrAjyXs8z5wKXA9UA/oAawEhhfaJxN4tJTnCgLuPe1KA8jMQoGvgS7Ar4C6wABgP9C30K6+vMZmwLV+KBMzC/LHcUXEQyFMxEfOuT3A53jCGABmVt3MnvW2Zuw1s3+aWU3vfQ3N7D9mdsjMDpjZwuMtFt5WjhHe7/ua2QozO+I9xt+9h//O+/WQt6VkgJm1MbNvzGy/maWZ2QwzCy9UT6KZ/a+ZrTOzw2Y2y8xqFLr/MjNb432uBDP7lXd7PTN73cx2m9lOb2tKkX+Ava/5BTPb5f33gndbe2BLoZq/KeKxI4Dzgcucc8udc3nOucPOuX84514vtOuLwHVm1qaEH8nfgP8t/PoLPY+Z2fPelqQj3tairkXsd56ZrS90+0szW17o9kIz+3Wh93aE9z2bCFzj/bmsLXTIlmb2g5mlm9kXx1u1inATEA1c7pzb6JwrcM7tc879n3PuU19eYyHPAH/2pcXUzIaaWYqZTfR+fhLN7IZC908zs1fN7FMzywTO8342pptZqpntMLNHzKyamXUC/gkM8L4Ph7zHKPZ3wnv/Lz6DZvYEMBh42Xusl737OjNr6/2+yDq8991sZt97n/egmf1kZiNLez9EAk0hTMRHZtYCGAnEF9r8FNAeTzBrCzQHHvPe9wcgBYjE00I0EShqnbBJwCTnXF2gDTDbu32I92u4t1tmMWDAk3haPzoBUcCfTjre1XhaV1oB3YGbvfX3BaYDE4Bw7/ETvY+ZBuR5X0MscAFQ3Pimh4H+3tfcA0/LzSPOua14WnaO1zysiMeOAJY555KLOfZxO4HXgD+XsM8KYAHwv0XcdwGe19ceT2vb1XhamU62BGhnnsAcguf9amZmYd7gEAcsLPwA59w84K/ALO/PpUehu68HxgKNgNBiagPP+zDPOZdRwusr7TUe9wFwBO/P2QdNgIZ4PqtjgMlm1qHQ/dcDTwBhwPfAS3jew9bAuXgC5Fjn3CbgTmCx9304HhSL/Z0o7jPonHsYz/v8O++xfldE3UXWUej+fnj+E9AQTzB93czMx/dEJCAUwkRK928zSweSgX3AH8HT2gLcDvzeOXfAOZeO54/z8a6hXKAp0NI5l+ucW+iKXqw1F2hrZg2dcxnOuSXFFeKci3fOfemcy3bOpeIZf3XuSbu96Jzb5Zw7AHzMf1vubgHe8D6+wDm30zm32cwaAxcB451zmc65fcDzFN/FdQPwuLflJhVPULqxuJpPEgHs9nHfJ4FRZtalhH0eA+42s8iTtufiCREdAXPObXLO/eJ5nXPHgOV4wkBvYC3wAzAIT9DcVtS4thJMdc5t9R53NoVaTU9yKu9Dca/xOIen6/ZR83Rz+uJR72foW+ATPCH1uA+dcz845wrwvI/XAg8559Kdc4nAcxTz8/bhd6LIz2BpxXpbZUurY4dz7jXnXD7wJp7fvca+vR0igaEQJlK6XzvnwoCheP6oH+9iigRqASvN0+V4CJjn3Q6erqR44Asz225mDxZz/FvwtBxsNrPlZnZJcYWYWWMze9fbZXgEeLtQPcftKfT9UeD4hQRRQEIRh20JhAC7C72Of+FpzSlKM2BHods7vNt8sR/PH8dSeQPey8DjJeyzAfgP8OBJ27/xPvYfwD4zm2xmdYs5zLd4frZDvN8vwBNsz/XePhXFvfcnO5X3ocjXeNI+n+Jpdb3Dh0MedM5lFrp98s+vcCtlQzyfjZN/3s2LOXZpvxPFfQZL40sdJ95759xR77dnfBGNiD8phIn4yNtqMA141rspDTgGdHHOhXv/1fMO4sf7P/Y/OOda4xmIfp+ZDS/iuNucc9fhCT1PA++b54q4olrN/urd3s3bfTkaTxelL5LxdHcWtT0baFjoddR1zhXXArULT3A7Ltq7zRdfAX29Xbu++BtwHp5WquL8EbiNk4KBc+5F51xvoDOekDuhmMefHMK+pfQQVtTP5lR8BVxoha58LEWRr/EkD+Pp8q5VyrHqn/S8J//8Cr+2NDytYSf/vHcWse/x/Yv9naD4z2BRxzr5uCXVIVIhKYSJnJoXgPPNrIe3u+Y14HkzawRgZs3N7ELv95eYWVtvF81hIB8oOPmAZjbazCK9xzvk3VwApHq/Fp4nKQzIAA6bWXOKDxZFeR0Ya2bDvQOrm5tZR2833RfAc+aZNqGaeS4AOLmb87h3gEfMLNI78PwxPC1ypXLOfQV8Ccw1s95mFuwdf3WnmY0rYv9DeLqd7i/hmPHALOCe49vMrI+Z9fOO88oEsijivfdaBHTAM7ZtmXPuRzx/7Pvx34sjTrYXiLFCU0OcorfwBJI5ZtbR+55HeAfMX3TyzkW9xiL2WQBswDPOqzR/Ns9UIYOBS4D3ijlmPp5u1Se8P6eWwH389+e9F2hxvBu0tN8JivkMFjpWkXOC+VCHSIWkECZyCrxdZNP57+D7B/B0OS7xdg9+hecPOkA77+0MYDHwinNufhGH/RXwo5ll4Bmkf61z7pi3S+UJ4Adv105/POOveuEJdZ/gGZTta+3L8Axkft77+G/5b8vCTXgGkm8EDuKZRqK47rK/4Bkwvg5YD6zybvPVlcCneELFYTzBIQ7Pe1WUSXgCbEkeBwq37tTFEwYO4um22o+nVe0XvF1zq4AfnXM53s2L8Ywx2lfM8x0PLfvNbFUptRX1nNl4BudvxhNKjwDL8HS7LS3mYSe/xqI8AjQoZZ89eN6XXcAM4M5SxmXdjSfIbsczUH8m8Ib3vm/wTEeyx8zSvNuK/Z0o5TM4CbjSe3Xji6dYh0iFZEWPExYRkcrGPDP+v+2c87U7WET8SC1hIiIiIgGgECYiIiISAOqOFBEREQkAtYSJiIiIBECpa42VBw0bNnQxMTGBLkNEREQqGOdgf2Y2qenZ5BU4WoTXpH5tXxeXOD0rV65Mc84Vt8rFCRUihMXExLBixYpAlyEiIiIVRFZuPjOXJvHqtwmQns2lrSO4d0Q7+reO8Ptzm9mO0veqICFMRERExBdZufm8syyJVxcksC89m36tGvDitbEMaOP/8HWqFMJERESkwsvKzefdZZ6Wr71HsunbqgGTymn4Ok4hTERERCqsrNx8Zi1P5pUF8Z7wFdOA56/pyYDWEXhWjSu/KmwIy83NJSUlhaysrECXInJCjRo1aNGiBSEhIYEuRUSkUsvKzWf2imRemZ/AniNZ9Impz/NX92RAm/Ifvo6rsCEsJSWFsLAwYmJiKsybLZWbc479+/eTkpJCq1atAl2OiEillJ2Xz+zlyfzDG77iWtbnuat7MLACha/jKmwIy8rKUgCTcsXMiIiIIDU1NdCliIhUOtl5+cxekcIr8+PZfTiL3i3r8+xVPRjUtuKFr+MqbAgDKuybLpWXPpMiImUrOy+f97zha9fhLHpFh/PMld05p23DCn/OrdAhTERERCqnnLwC3luZzD++8YSv2OhwnvpNdwa3q/jh6zgtW3QG9uzZw7XXXkubNm3o3bs3F110EVu3bi12/8TERLp27QrAggULuOSSS0o8/po1a/j000/LtOY6der8YtuhQ4d45ZVXTvuYMTExpKWlnUlZ3HrrrWzcuBGAv/71rye2F37PSjN9+nS6du1Kt27diI2N5dlnnwXg5ptvpnnz5mRnZwOQlpbG8RUYEhMTMTNeeumlE8f53e9+x7Rp035x/OzsbK655hratm1Lv379SExMPI1XKiIiJcnJK2Dm0iTOe3YBD8/dQKO6NXhzXF8++O1AhrSPrDQBDBTCTptzjssvv5yhQ4eSkJDAypUrefLJJ9m7d2+ZPcfphLC8vLxTfp4zDWFlYcqUKXTu3Bn4eQjz1WeffcYLL7zAF198wfr161myZAn16tU7cX9QUBBvvPFGkY9t1KgRkyZNIicnp8TneP3116lfvz7x8fH8/ve/54EHHjjlOkVEpGg5eQW8s8wTvibOXU9kWHXeHNeXuf8zkHMrWfg6TiHsNM2fP5+QkBDuvPPOE9t69OjB4MGDcc4xYcKEE60ys2bNKvFYy5YtY8CAAcTGxjJw4EC2bNlCTk4Ojz32GLNmzaJnz57MmjWLzMxMxo0bR9++fYmNjeXDDz8EYNq0aVx66aUMGzaM4cOHk5GRwfDhw+nVqxfdunU7sV9xHnzwQRISEujZsycTJkwo9vGZmZlcfPHF9OjRg65du/7idR07doyRI0fy2muv/Wz7e++9x3333QfApEmTaN26NQDbt29n0KBBAAwdOpQVK1bw4IMPcuzYMXr27MkNN9wAQH5+PrfddhtdunThggsu4NixY794DU8++STPPvsszZo1A6B69ercdtttJ+4fP348zz//fJEhNTIykuHDh/Pmm2+W+D59+OGHjBkzBoArr7ySr7/+GudciY8REZGS5eb/N3w99MF6GoZVZ9rYPpU6fB3ntzFhZtYBKPxXujXwGDDduz0GSASuds4dPJPn+vPHP7Jx15EzOcQvdG5Wlz+O6lLs/Rs2bKB3795F3vfBBx+wZs0a1q5dS1paGn369GHIkCHFHqtjx44sXLiQ4OBgvvrqKyZOnMicOXN4/PHHWbFiBS+//DIAEydOZNiwYbzxxhscOnSIvn37MmLECABWrVrFunXraNCgAXl5ecydO5e6deuSlpZG//79ufTSS4v9ID/11FNs2LCBNWvWABT7+Hnz5tGsWTM++eQTAA4fPnziGBkZGVx77bXcdNNN3HTTTT87/uDBg3nmmWcAWLhwIREREezcuZOFCxf+4n156qmnePnll0/UkpiYyLZt23jnnXd47bXXuPrqq5kzZw6jR4/2+ecBEB0dzTnnnMNbb73FqFGjfnH/Aw88wMiRIxk3blyxx9i5cydRUVEABAcHU69ePfbv30/Dhg2LfYyIVHzOOb6PT+Pd5cmEBlVjQJsIBrVtSPPwmoEurULLzS9gzsoUXp4fT8rBY/RoUY+/XN6VoZU8eBXmtxDmnNsC9AQwsyBgJzAXeBD42jn3lJk96L1dqfp1vv/+e6677jqCgoJo3Lgx5557LsuXL6d79+5F7n/48GHGjBnDtm3bMDNyc3OL3O+LL77go48+OjHWKSsri6SkJADOP/98GjRoAHhOGBMnTuS7776jWrVq7Ny5k71799KkSROf6i/u8d26deMPf/gDDzzwAJdccgmDBw8+8ZjLLruM+++//0TrVWFNmjQhIyOD9PR0kpOTuf766/nuu+9YuHAhV1xxRan1tGrVip49ewLQu3fv0x6L9dBDD3HZZZdx8cUX/+K+1q1b069fP2bOnHlaxxaRyic9K5c5K1OYvmQH21MziagdihnMXb0TgJYRtRjYpiGD2kYwoHUEEXWqB7jiiiE3v4APVqXw0jee8NW9RT3+77KuDO1QdcLXcWfr6sjhQIJzboeZXQYM9W5/E1jAGYawklqs/KVLly68//77ZXKsRx99lPPOO4+5c+eSmJjI0KFDi9zPOcecOXPo0KHDz7YvXbqU2rVrn7g9Y8YMUlNTWblyJSEhIcTExJzSygLFPb59+/asWrWKTz/9lEceeYThw4fz2GOPATBo0CDmzZvH9ddfX+Qv0cCBA5k6dSodOnRg8ODBvPHGGyxevJjnnnuu1HqqV//viS0oKKjI7sguXbqwcuVKhg0bVuxx2rVrR8+ePZk9e3aR90+cOJErr7ySc889t8j7mzdvTnJyMi1atCAvL4/Dhw8TEVF+1yQTkdMTvy+D6YsTmbMyhcycfHpEhfP8NT24qFtTQoOqsXVvBj/Ep7EoIY3/rN3FO8s8/xnu2CSMQW0bMrBNBH1bNSCshlbOKCw3v4C5q3by0vxtJB/whK/HL+vCeR0aVbnwddzZGhN2LfCO9/vGzrnd3u/3AI2LeoCZ3W5mK8xsRXmc/HLYsGFkZ2czefLkE9vWrVvHwoULGTx4MLNmzSI/P5/U1FS+++47+vbtW+yxDh8+TPPmzQF+dlVeWFgY6enpJ25feOGFvPTSSyfGIa1evbrY4zVq1IiQkBDmz5/Pjh07SnwtJz9PcY/ftWsXtWrVYvTo0UyYMIFVq1adeMzjjz9O/fr1ueuuu4p8jsGDB/Pss88yZMgQYmNjmT9/PtWrV//Z4PnjQkJCim0NLM5DDz3EhAkT2LNnDwA5OTlMmTLlF/s9/PDDJ1oST9axY0c6d+7Mxx9/XOT9l1566YlxY++//z7Dhg2rsicOkcomv8DxxY97GD1lKSP+/i3vLkvmwq5N+Pddg/jwrkFcHtuC6sFBmBkdmoQx7pxWTBnTh9WPnc/c/xnIhAs7EFEnlLeW7OCWN1fQ8/EvueKVH3j28y0sSkgjKzc/0C8xYPLyC5i9Ipnhz33L/XPWEV4zlNfHxPHhXYMY1rFxlT6P+r0lzMxCgUuBh06+zznnzKzIkc3OucnAZIC4uLhyN/rZzJg7dy7jx4/n6aefpkaNGsTExPDCCy9wzjnnsHjxYnr06IGZ8cwzz9CkSZNiu9Huv/9+xowZw1/+8pefdZWdd955PPXUU/Ts2ZOHHnqIRx99lPHjx9O9e3cKCgpo1aoV//nPf35xvBtuuIFRo0bRrVs34uLi6NixY4mvJSIigkGDBtG1a1dGjhzJAw88UOTj169fz4QJE6hWrRohISG8+uqrPzvOpEmTGDduHPfff/+JMWDHDR48mOTkZIYMGUJQUBBRUVHF1nX77bfTvXt3evXqxRNPPFFi7cdddNFF7N27lxEjRuCcw8yKHN/VpUsXevXq9bMAWdjDDz9MbGxskffdcsst3HjjjbRt25YGDRrw7rvv+lSbiJRfBzNzmLUimbcW72DnoWM0rVeDCRd24Jo+UTT0oXsxOKgasdH1iY2uz13ntSUrN59VOw6yKGE/PySk8eq3Cbw8P57qwdWIi6nv7b5sSNdmdQkOqtzXxuXlFzB39U5enh/Pjv1H6dq8LlNuimN4p6rb8nUy8/fVXd7ux7uccxd4b28BhjrndptZU2CBc65DSceIi4tzK1as+Nm2TZs20alTJ3+VLXLa9NkUKf827DzM9MWJfLhmF9l5BfRv3YAxA2I4v3PjMg1H6Vm5LPvpAD/E72dRQhqb93h6HcKqB9OvdQQDvYP82zeuU2mCSV5+Af9es4uXvtnGjv1H6dKsLuNHtGdEFQpfZrbSORdX2n5nY0zYdfy3KxLgI2AM8JT3a8nzJ4iIiJSB3PwCPtuwh+mLElmx4yA1Q4L4Te8W3DSgJR2b1PXLc4bVCGF4p8YM7+QZeZOWkc3ihP0sSvCEsq82eeaWbFgnlAFtGjLIG8qiGtTySz3+lJdfwIfe8JW4/yidm9Zl8o29Ob9z1e5yLIlfQ5iZ1QbOB+4otPkpYLaZ3QLsAK72Zw0iIlK17UvPYubSJGYuTWJfejYtI2rxyMWduCouino1z+7g+YZ1qjOqRzNG9fDMaZhy8KgnkMWn8UPCfj5euwuAFvVrMqhNQwa2jWBAmwgahdU4q3Weirz8Aj5au4uXvonnp7RMOjWty79u7M0FCl+l8msIc85lAhEnbduP52rJsji+fsBSrmjyVpHywTnHqqSDvLloB59t2E1uvmNoh0ieHhDDue0jqVatfPztaFG/FlfH1eLquCiccySkZvBD/H5+iE/jsw27mbUiGYD2jeswsI3nyst+rSPOengsSn6B46O1O3np63i2p2XSsUkY/xztCV/l5f0t7/w+JqwsFDUm7KeffiIsLIyIiAgFMSkXnHPs37+f9PR0WrVqFehyRKqkrNx8Plq7i+mLE9mw8whhNYK5qncUNw5oSauGtUt9fHmSX+D4cddhzyD/+DSWJx4gK7eAagbdmtdjYNuGDGrTkLiY+tQICTqrdX28dhcvfr3tRPgaP6IdF3RuovDl5euYsAobwnJzc0lJSTml+a9E/K1GjRq0aNGCkJDA/y9VpCpJOXiUt5ckMWt5EgeP5tK+cR1uGhDD5bHNqV39bE2J6V/ZefmsSTrED97uyzXJh8grcIQGVaNXy/AT3ZfdW4QT4ocrL/MLHP9Zt4tJX29je6onfN07vB0XdlH4OlmlD2EiIlK1OedYlLCfNxclnhjgfkHnJowZGEP/1g0qfS9JZnYeyxIPsCg+jUUJ+9m4+wjOQe3QIPq2auCdOLYhHZuEnVFIOjl8dWgcxr0j2vErha9ilaerI0VERMpMRnYec1el8ObiHcTvy6BB7VB+O7QN1/drWaXWc6xdPZjzOjTivA6NAM+cZ4u3e666XBS/n/lbNgHQoHYoA1pHMLBtBAPbNCQmopZPAfV4+Hrx620keMPXKzf0UvgqQ2oJExGRCmF7agbTF+9gzsoU0rPz6N6iHmMGxHBx96ZndUxURbH78DEWxXsmjV0Uv589RzzDd5rVq8FA7/JKg9o2pHHdn195mV/g+GT9bl78ehvx+zJo37gO9w5vz8iuCl++UnekiIhUePkFjgVb9vHm4h18tzWVkCDjku7NuGlAS3pGhVf6Lsey4pzjp7RMfkjYz+IET/floaOe5eHaRNY+sRB5dl4BL38Tz7Z9GbRrVId7R7Tjoq5NFb5OkUKYiIhUWIeP5jJ7RTJvLdlB0oGjNK5bndH9WnJt32giw0pfTkhKVlDg2Lj7CIu9yyst++kAR3M861u2bVSHe4e346JuTQlS+DotGhMmIiIVzqbdR5i+OJG5q3eSlVtA31YNeOBXHbmgS2O/XPFXVVWrZnRtXo+uzetx25DW5OQVsC7lEOnZeQxpF6nwdZYohImISEDl5hfwxY97eXNRIssSD1AjpBqXxzbnxv4xdG7mn+WE5OdCg6sRF9Mg0GVUOQphIiISEKnp2by7LIkZS5PYcySLqAY1efiiTlwV14LwWqGBLk/E7xTCRETkrHHOsSb5EG8uSuST9Z7lhIa0j+SJy7sytEMjdYNJlaIQJiIifpeVm88n63bz5uJE1qUcpk71YG7o15KbBrSkdWSdQJcnEhAKYSIi4je7Dh3j7SU7eHd5Mgcyc2jbqA7/d1kXLu/VgjqVZDkhkdOl3wARESlTzjmWbD/Am4sS+WLjHgBGdGrMmIExDGwTobm9RLwUwkREpExkZufx7zU7mb5oB1v2phNeK4Tbh7RhdP9oWtSvFejyRModhTARETljm/cc4Zp/LeHwsVy6NKvLM1d259IezbSckEgJFMJEROSMPfHJJsxgzm8H0Cu6vrocRXyg6YdFROSMfLc1lYXb0rh7WDt6t2ygACbiI4UwERE5bfkFjic/20xUg5qM7h8d6HJEKhSFMBEROW3/Xr2TTbuPMOHCjlQP1vgvkVOhECYiIqclKzef577YQvcW9bikW9NAlyNS4SiEiYjIaZm2KJFdh7N4cGRHqmm5IZFTphAmIiKn7GBmDv+YH8+wjo0Y2KZhoMsRqZAUwkRE5JS9PD+ezOw8HvhVx0CXIlJhKYSJiMgpST5wlOmLE7mqdxQdmoQFuhyRCkshTERETsnfPt9CUDXj9+e3D3QpIhWaX0OYmYWb2ftmttnMNpnZADPraWZLzGyNma0ws77+rEFERMrOupRDfLR2F7ee05om9WoEuhyRCs3fyxZNAuY55640s1CgFjAb+LNz7jMzuwh4Bhjq5zpEROQMOef466ebaFA7lDvObR3ockQqPL+1hJlZPWAI8DqAcy7HOXcIcEBd7271gF3+qkFERMrOgi2pLNl+gHuHtyOsRkigyxGp8PzZEtYKSAWmmlkPYCVwLzAe+NzMnsUTAgf6sQYRESkDnuWJNhETUYvr+mp5IpGy4M8xYcFAL+BV51wskAk8CPwW+L1zLgr4Pd6WspOZ2e3eMWMrUlNT/VimiIiUZs7KFLbuzeD+X3UkNFjXdImUBX/+JqUAKc65pd7b7+MJZWOAD7zb3gOKHJjvnJvsnItzzsVFRkb6sUwRESnJsZx8nvtyCz2jwhnZtUmgyxGpNPwWwpxze4BkM+vg3TQc2IhnDNi53m3DgG3+qkFERM7cGz/8xN4j2Uy8qBNmWp5IpKz4++rIu4EZ3isjtwNjgQ+BSWYWDGQBt/u5BhEROU37M7J5dUEC53duTN9WDQJdjkil4tcQ5pxbA8SdtPl7oLc/n1dERMrGS9/Ecyw3X8sTifiBRleKiEiREtMyeXvJDq7pE0XbRnUCXY5IpaMQJiIiRfrb51sIDa7G+BHtAl2KSKWkECYiIr+wOukgn6zfzW2DW9MoTMsTifiDQpiIiPyMc44nP91MwzrVuW2IlicS8ReFMBER+ZmvNu1jWeIBxo9oR53q/r6IXqTqUggTEZET8vILeOqzTbSOrM01faICXY5IpaYQJiIiJ8xekUJCaiYP/KojIUH6EyHiT/oNExERADKz83j+q63EtazPBZ0bB7ockUpPIUxERACYsvAnUtOzeUjLE4mcFQphIiJCano2//ougZFdm9C7Zf1AlyNSJSiEiYgIk77eSk5eARMu7BDXiq7OAAAgAElEQVToUkSqDIUwEZEqLiE1g3eWJXN9v2haR2p5IpGzRSFMRKSKe2beZmoEV+Oe4VqeSORsUggTEanCViQe4PMf93LnuW1oWKd6oMsRqVIUwkREqijnHH/9dBONwqpzy+BWgS5HpMpRCBMRqaI+/3EPq5IOcd/57akVquWJRM42hTARkSooN7+Ap+dtoV2jOlzZu0WgyxGpkhTCRESqoHeXJfFTWiYPjuxIsJYnEgmIU/rNM7M2ZtbNX8WIiIj/ZWTn8cJX2+jXqgHDOjYKdDkiVZbPgwDMbCLQFigws+rOuRv9V5aIiPjL5G8T2J+Zw+tankgkoIoNYWZ2D/AP51y+d1MP59w13vvWnY3iRESkbO09ksVrC3/iku5N6RkVHuhyRKq0kroj9wPzzOxS7+0vzGyemX0BfO7/0kREpKy98NVW8gq0PJFIeVBsCHPOzQBGAd3N7CNgJXAFcJVzbsJZqk9ERMrItr3pzFqezOj+LWkZUTvQ5YhUeaUNzG8DzAZuB+4CJgE1/V2UiIiUvafnbaZ2aDB3D9PyRCLlQUljwqYBuUAtYKdz7jYziwVeM7PlzrnHz1KNIiJyhpZs389Xm/Zx/6860KB2aKDLERFKvjoy1jnXA8DMVgM451YDo8zssrNRnIiInDnnHE9+uomm9WowbpCWJxIpL0rqjvzMzD43s2+AmYXvcM596N+yRKQqyckrYOvedHLyCgJdSqX0yfrdrE05zH3nt6dGSFCgyxERr2JbwpxzD5pZXaDAOZdxOgc3s3BgCtAVcMA459xiM7sbzxizfOAT59z9p3N8EamYdh8+xuqkQ6xOOsjqpEOs33mY7LwCeresz5vj+lKnutYxLCs5eQU8M28LHZuEcUUvLU8kUp6UeKZzzh05w+NPAuY55640s1CglpmdB1yGZ96xbDPTdM0ilVhWbj4bdh72hK7kg6zacYg9R7IACA2uRrfm9bixf0vq1w7l719uZcwbyxTEytCMpTtIOnCUaWP7EFRNE7OKlCd+O8uZWT1gCHAzgHMuB8gxs98CTznnsr3b9/mrBhE5u5xzJB84xupkTwvXqqSDbNx1hLwCB0BUg5r0bdWA2OhwYqPr06lpGNWD/9s91rphbX73zmoFsTJyJCuXF7/exqC2EZzbPjLQ5YjISfx5hmsFpAJTzawHnnnG7gXaA4PN7AkgC/hf59zykx9sZrfjmRqD6OhoP5YpIqcrMzuPtSmHfta1uD8zB4CaIUH0iKrHbUNaExsVTs/ocBqF1SjxeCO7NeVlUBArI/9ckMDBo7k8NFLLE4mURz6d3cxsIBBTeH/n3HQfjt0LuNs5t9TMJgEPerc3APoDfYDZZtbaOecKP9g5NxmYDBAXF/ez+0Tk7CsocGxPy2R10kFWeUPX1r3peBu5aB1Zm6EdGnlbucLp0DiM4KDSpiL8JQWxsrH78DFe//4nft2zGV2b1wt0OSJShFLPbGb2Fp5JW9fgGUgPnkH2pYWwFCDFObfUe/t9PCEsBfjAG7qWmVkB0BBPq5mIlBOHj+ayJuUQq3YcZHXyIdYkHeRIVh4AYTWC6RkVzgVdmhAbHU7PFuHUL8O5pxTEztzfv9iKc/CHC7Q8kUh55ctZLQ7ofHJLVWmcc3vMLNnMOjjntgDDgY1AAnAeMN/M2gOhQNop1i0iZSi/wLF1b/qJcVyrkw6SkJoJgBl0aBzGxd2bEhtVn9jocNpE1qGanwd5K4idvs17jvD+qhRuPacVUQ1qBbocESmGL2e0DUATYPdpHP9uYIb3ysjtwFggE3jDzDYAOcCYUw14InJm0jKyWXMicB1iXcohMnM8Dd0NaocSGxXO5bHNiY2uT/cW9QirERKQOhXETs9Tn20mrHowd53XNtCliEgJfDmbNQQ2mtkyIPv4RufcpaU90Dm3Bk9L2slG+1yhiJyRnLwCNu85cqJbcXXSIZIOHAUgqJrRuWldftO7hWcsV1R9WkbUKleDuAsHsZvfWMY0BbES/RCfxoItqUy8qCPhtbQ8kUh55suZ7E/+LkJEys6ew1newfM/nwgVoFFYdXpF1+eGftHERtenW/N61Awt/zOoK4j5pqDA8eRnm2geXpObBsQEuhwRKUWpZzHn3LdnoxAROXVZufn8uOswq3YcOjE31+7D3olQg6rRtXldRvdvSWx0OL2i69O0Xo1y1cp1KkZ2a8pLwN0KYsX6eN0uNuw8wvPX9NDyRCIVQLFnMDP73jl3jpml47ka8sRdgHPO1fV7dSJygnOOlIPHTrRwrU46yMbdR8jN9/x6tqhfk7iYBsRGhdOr5S8nQq0MLurWFFAQK0p2Xj7PzNtC56Z1uaxH80CXIyI+KGntyHO8X8POXjkicrL4fenMWJrEx2t3k5bhGZZZMySI7i3qccs5rU/My1XaRKiVhYJY0d5avIOdh47x9G+6+/3KVREpGzpziZRD2Xn5zNuwhxlLk1j20wFCgozzOzdmQJuG9DqDiVArCwWxnzt8NJeXvolnSPtIzmnXMNDliIiPqu5ZS6QcSkzL5J1lSby3MoUDmTlEN6jFgyM7cmXvFjSsUz3Q5ZUrCmL/9cqCeI5k5fLgrzoGuhQROQVV84wlUo7k5hfw5ca9zFyaxPfxaQRVM87v1Jgb+kczqE1DdS2VQEEMUg4eZeqiRK6IbUHnZhqqK1KR+LJs0d3A2865g2ehHpEqI/nAUd5dnsTsFSmkpmfTPLwmfzi/PVf3iaJx3aoxvqssFA5iY6cuY+rYqhXE/v7FVgD+cEH7AFciIqfKlzNVY2C5ma0C3gA+1wz3IqcnL7+A+VtSmbl0Bwu2pmLAsI6NuL5fNOe2b0SQWr1OS1UNYht2Hmbump3cMaQNzcJrBrocETlFvswT9oiZPQpcgGfZoZfNbDbwunMuwd8FilQGew5n8e7yJGYtT2b34SwahVXn7vPack3faJrrj2eZuKhbU5yDe96tOkHs6XmbqVczhN8ObRPoUkTkNPh0hnLOOTPbA+wB8oD6wPtm9qVz7n5/FihSURUUOL7blsrMpUl8vXkf+QWOIe0j+eOoLgzv1IiQKnx1o79c3N3TIlYVgth3W1NZuC2NRy/pTL2agVnbU0TOjC9jwu4FbgLSgCnABOdcrplVA7YBCmEihaSmZzN7RTLvLk8i+cAxImqHctvg1lzXN4qWEbUDXV6lVxWCWH6B48nPNhPVoCaj+0cHuhwROU2+nJkaAFc453YU3uicKzCzS/xTlkjF4pxjccJ+ZixN4vMf95BX4BjQOoL7L+zIhV2aEBqsVq+zqbIHsX+v3smm3Ud48brYSrcqgkhV4stZ6TPgwPEbZlYX6OScW+qc2+S3ykQqgIOZOby/MoWZy5L4KS2T8Foh3Dwwhuv6RdMmsk6gy6vSKmsQy8rN57kvttC9RT0u8V6QICIVky9npFeBXoVuZxSxTaTKcM6xYsdBZizZwacb9pCTV0Bcy/rcM7wtI7s21cLJ5UhlDGLTFiWy63AWz17dQ3PIiVRwvpyNrPCUFN5uyIp9FhM5DYeP5TJ3VQozliaxbV8GYdWDua5PFNf3a0mHJlpitbw6OYhNG9uX2hU0iB3MzOEf8+MZ1rERA9toeSKRis6XM9F2M7sHT+sXwP8A2/1Xkkj54ZxjTfIhZi5N4uN1u8jKLaBHi3o885vuXNKjKbVCK+Yf86qmcBC7uQIHsZfnx5OZnccDWp5IpFLw5Sx0J/Ai8AjggK+B2/1ZlEigZWTn8eGancxYksTG3UeoFRrE5bEtuKFfNF2b1wt0eXIaLu7eFIfj3nfXVMgglnzgKNMXJ3JV7yi1vIpUEr5M1roPuPYs1CIScBt2HmbmsiQ+XL2TzJx8OjWty19+3ZXLejYjrIbmYqroLuneDKBCBrG/fb6FoGrG78/X8kQilYUv84TVAG4BugAnFrRzzo3zY10iZ82xnHw+XreLGUuTWJt8iOrB1RjVoxk39IumZ1Q4Zhr8XJlUxCC2LuUQH63dxe/Oa0uTelpXVKSy8OXM8xawGbgQeBy4AdDUFFLhbd2bzsylScxZlUJ6Vh5tG9Xhj6M6c0VsC+rVUqtXZVaRgphzjr9+uokGtUO549zWgS5HRMqQL2edts65q8zsMufcm2Y2E1jo78JE/CErN595G/YwY+kOliceJDSoGiO7NeGGfi3pE1NfrV5VSEUJYgu2pLJk+wH+fGkXdYmLVDK+nHFyvV8PmVlXPOtHNvJfSSJlb3tqBu8sS+L9lSkcPJpLTEQtJl7UkSt7R9Ggdmigy5MAKe9BzLM80SZiImpxXV8tTyRS2fhytplsZvXxXB35EVAHeNSvVYmUgZy8Ar7YuIeZS5NYlLCf4GrGhV2acH2/aAa0jtBElwL8PIiNnbqcqWP7lJsgNmdlClv3ZvDKDb209JVIJVTimca7SPcR59xB4DtAAxKk3Es+cJSZy5J4b0UyaRk5NA+vyYQLO3BVXAsahWlQs/xSeQxix3Lyee7LLfSMCmdk1yYBrUVE/KPEs4x3dvz7gdlnqR6R05KXX8DXm/cxc2kS321LxYDhnRpzfb9ohrSLJEitXlKKS7o3wzkYP6t8BLE3fviJvUeyeem6XhqrKFJJ+XKG+crM/heYBWQe3+icO1D8QzzMLByYAnTFM9HrOOfcYu99fwCeBSKdc2mnUbsI4FnKZey05axJPkSTujW4Z1g7ru0bRdN6NQNdmlQwo3p4WsQCHcT2Z2Tz6oIEzu/cmL6tGpz15xeRs8OXs8s13q93Fdrm8K1rchIwzzl3pZmFArUAzCwKuABIOoVaRX5h35Esbnx9GT/tz+S5q3pwWc9mBAdp7IycvvIQxF76Jp5juflankikkvNlxvxWp3NgM6sHDAFu9h4nB8jx3v08cD/w4ekcWwQg5eBRRk9Zyr70bKbd3IeBbbWgsZSNQAaxxLRM3l6yg2v6RNG2UZ2z8pwiEhi+zJh/U1HbnXPTS3loKyAVmGpmPYCVwL3ACGCnc25tSeMczOx2vGtURkfr0mz5uYTUDEZPWUpmdh5v39qPXtH1A12SVDKBCmJ/+3wLocHVGD+ind+fS0QCy5d+mz6F/g0G/gRc6sPjgoFewKvOuVg848n+BEwEHivtwc65yc65OOdcXGRkpA9PJ1XFxl1HuPqfi8nNL+Dd2wcogInfjOrRjBeu6cnKpIOMnbqczOw8vz7f6qSDfLJ+N7cNbq0reUWqAF+6I+8ufNs72P5dH46dAqQ455Z6b7+PJ4S1Ao63grUAVplZX+fcnlOoW6qolTsOMnbqMmpXD+btW/vRJlLdNeJfP2sRm7acqTf7p0XMOceTn26mYZ3q3DZEswGJVAWnM4I5E0+QKpE3VCWbWQfvpuHAKudcI+dcjHMuBk9Q66UAJr74IT6NG19fSoPaobx35wAFMDlrTrSI7TjI2Gn+aRH7atM+liUeYPyIdtQpJ5PFioh/+TIm7GM8V0OCJ7R1xvd5w+4GZnivjNwOjD2dIkW+3LiXu2asonVkbabf0lddNXLWHW8Ru/fd1WXeIpaXX8BTn22idWRtrukTVSbHFJHyz5czyLOFvs8DdjjnUnw5uHNuDRBXwv0xvhxHqrYP1+zkvtlr6dq8Hm+O7UN4La31KIHhryA2e0UKCamZ/OvG3oRoihWRKsOX3/YkYKlz7lvn3A/AfjOL8WtVIl4zlyYxftYa+sTUZ8at/RTAJOBG9WjGpGtjWZF4oEy6JjOz83j+q63EtazPBZ0bl1GVIlIR+BLC3gMKCt3O924T8avJ3yUwce56zuvQiGlj+2qcjJQbZRnEpiz8idT0bB66qJOWJxKpYnwJYcHeiVaBE5OuqjlC/MY5x9+/2MJfP93Mxd2b8s/RvakREhToskR+piyCWGp6Nv/6LoGRXZvQu6WmWhGpanwJYalmdmJeMDO7DNBaj+IXBQWOP3+8kRe/iefaPlG8eG0socEaIyPl05kGsUlfbyUnr4AJF3YofWcRqXR8+et2JzDRzJLMLAl4ALjDv2VJVZRf4HhgzjqmLUrklnNa8eQV3Qiqpu4ZKd9ODmJHc3wLYgmpGbyzLJnr+0XTWtOtiFRJvkzWmgD0N7M63tsZfq9KqpycvAJ+P2sNn6zfzfgR7bh3eDuNj5EKo/BVkzdPXc60sX2oFVry6fWZeZupEVyNe4ZreSKRqqrUljAz+6uZhTvnMpxzGWZW38z+cjaKk6rhWE4+t7+1gk/W7+aRizsxfkR7BTCpcEb1aMYL3haxm6eW3CK2IvEAn/+4lzvPbUPDOtXPYpUiUp740h050jl36PgN59xB4CL/lSRVSXpWLmOmLuPbrak8eUU3bh2s5Vqk4rrUhyDmnOOvn26iUVh1bhlc6uIjIlKJ+RLCgszsxH/VzKwmoP+6yRk7mJnDDVOWsmrHQSZdG8t1faMDXZLIGSstiH3+4x5WJR3ivvPbl9plKSKVmy8hbAbwtZndYma3AF8C0/1bllR2+45kcc3kxWzek86/buzNpd4xNSKVQXFBLDe/gKfnbaFdozpc2btFgKsUkUDzZWD+02a2Fhjh3fR/zrnP/VuWVGbJB44y+vWlpKVnM21sHwa2aRjokkTK3PH/WIwvNFh/zsoUfkrL5PUxcQRreSKRKs+ntnDn3DxgHoCZnWNm/3DO3eXXyqRSit+XwegpSzmWm8/bt/YjNloTVErldXIQS9iXQb9WDRjWsVGAKxOR8sCnEGZmscB1wNXAT8AH/ixKKqcNOw8z5o1lmBnv3t6fTk3rBrokEb8rHMQKHLyu5YlExKvYEGZm7fEEr+vwzJA/CzDn3HlnqTapRFbu8IyNCasezNu39tPklFKlXNqjGXWqB7Hz4DF6RoUHuhwRKSdKagnbDCwELnHOxQOY2e/PSlVSqXy/LY3bpq+gSb0avH1rP5qH1wx0SSJn3bCOjQNdgoiUMyWNDL0C2A3MN7PXzGw4oDZ0OSVf/LiHcdOW0zKiFrPvGKAAJiIi4lVsCHPO/ds5dy3QEZgPjAcamdmrZnbB2SpQKq65q1P47YxVdG5Wl1m3DyAyTNPLiYiIHFfqNdLOuUzn3Ezn3CigBbAazyLeIsV6e8kO7pu9lr4xDXj71n7UqxUS6JJERETKlVOaqMY5d9A5N9k5N9xfBUnF989vE3jk3xsY1qERU8f2oU51zQouIiJyMv11lDLjnOO5L7by8vx4RvVoxt+v7kGIJqQUEREpkkKYlImCAsfj/9nItEWJXNsniicu70ZQNV3HISIiUhyFMDljefkFPDBnPXNWpXDb4FZM1GSUIiIipVIIkzOSnZfP+HfX8NmGPdx3fnvuHtZWAUxERMQHCmFy2o7l5HPH2yv5bmsqj17SmVvOaRXokkRERCoMhTA5LUeycrl12gqW7zjA07/pxjV9ogNdkoiISIWiECan7EBmDmPeWMam3Ud48dpYRnkXKBYRERHfKYTJKdl7JIvRU5aSdOAok2/qrfXwRERETpNfQ5iZhQNTgK6AA8bhWZNyFJADJABjnXOH/FmHlI3kA0e5YcpS9mdkM21sXwa0iQh0SSIiIhWWv2fSnATMc851BHoAm4Avga7Oue7AVuAhP9cgZSB+XzpX/nMRh4/lMuO2/gpgIiIiZ8hvIczM6gFDgNcBnHM5zrlDzrkvnHN53t2W4FmPUsqxDTsPc/W/lpBfALPu6E/PqPBAlyQiIlLh+bMlrBWQCkw1s9VmNsXMap+0zzjgs6IebGa3m9kKM1uRmprqxzKlJCsSD3Dd5CXUDAnivTsH0LFJ3UCXJCIiUin4M4QFA72AV51zsUAm8ODxO83sYSAPmFHUg70Lhcc55+IiIyP9WKYUZ+G2VG58fRmRYdV5784BtGp4coYWERGR0+XPEJYCpDjnlnpvv48nlGFmNwOXADc455wfa5DTNG/DHm6ZtoKYhrWZdccAmoXXDHRJIiIilYrfQphzbg+QbGYdvJuGAxvN7FfA/cClzrmj/np+OX0frErhrpmr6NK8Lu/e1p/IsOqBLklERKTS8fc8YXcDM8wsFNgOjAWWA9WBL71rDC5xzt3p5zrER28tTuTRD39kUNsIJt8YR+3qmkpORETEH/z6F9Y5twaIO2lzW38+p5y+VxbE88y8LYzo1JiXr4+lRkhQoEsSERGptNTMITjn+NvnW3hlQQKX9WzGs1f1ICTI31PIiYiIVG0KYVVcQYHjTx//yPTFO7iubzR/+XVXgqpZoMsSERGp9BTCqrC8/ALun7OOD1bt5PYhrXloZEe84/RERETEzxTCqqjsvHzueWc1n/+4lz+c357fDWurACYiInIWKYRVQUdz8rjjrZUs3JbGH0d1ZuygVoEuSUREpMpRCKtijmTlMm7qclYlHeSZK7tzdVxUoEsSERGpkhTCqpD9GdmMmbqMLXvSeem6XlzcvWmgSxIREamyFMKqiD2Hsxj9+lKSDxxl8k1xnNehUaBLEhERqdIUwqqALXvSueXN5Rw6msub4/rSv3VEoEsSERGp8hTCKrn/rNvFhPfWUadGMDNu7UePqPBAlyQiIiIohFVa+QWOZz7fzL++3U7vlvV59YZeNKpbI9BliYiIiJdCWCV0MDOHe95dzcJtaYzuH81jl3QhNFjLEImIiJQnCmGVzI+7DnPHWyvZdySbp3/TjWv6RAe6JBERESmCQlgl8uGanTwwZx3hNUOZdUd/YqPrB7okERERKYZCWCWQl1/AU59tZsr3P9E3pgH/uKEXkWHVA12WiIiIlEAhrILbn5HN3e+sZlHCfm4eGMPDF3ciJEjjv0RERMo7hbAKbMNOz/iv1Ixsnr2qB1f2bhHokkRERMRHCmEV1AerUnjog/VE1A5lzp0D6daiXqBLEhERkVOgEFbB5OYX8MQnm5i2KJH+rRvwj+t7EVFH479EREQqGoWwCiQ1PZu7Zq5i2U8HuOWcVjw0siPBGv8lIiJSISmEVRBrkw9x59srOXg0hxeu6cmvY5sHuiQRERE5AwphFcDs5ck88uEGGoVVZ85vB9KlmcZ/iYiIVHQKYeVYTl4Bj//nR95eksQ5bRvy0nWx1K8dGuiyREREpAwohJVT+45k8T8zVrFix0HuGNKaCRd20PgvERGRSkQhrBxaueMgv317JelZebx0XSyjejQLdEkiIiJSxhTCypmZS5P440cbaFqvJtNv6UvHJnUDXZKIiIj4gV9DmJmFA1OAroADxgFbgFlADJAIXO2cO+jPOiqC7Lx8/vTRj7yzLJkh7SN58dqehNfS+C8REZHKyt+DjCYB85xzHYEewCbgQeBr51w74Gvv7Sptz+EsrvnXEt5Zlsxd57Vh6s19FMBEREQqOb+1hJlZPWAIcDOAcy4HyDGzy4Ch3t3eBBYAD/irjvJueeIBfvv2Ko7m5PHqDb0Y2a1poEsSERGRs8CfLWGtgFRgqpmtNrMpZlYbaOyc2+3dZw/QuKgHm9ntZrbCzFakpqb6sczAcM7x1uJErpu8hLAawfz7rkEKYCIiIlWIP0NYMNALeNU5FwtkclLXo3PO4Rkr9gvOucnOuTjnXFxkZKQfyzz7snLzuf/9dTz64Y8MaR/Jv+8aRPvGYYEuS0RERM4ifw7MTwFSnHNLvbffxxPC9ppZU+fcbjNrCuzzYw3lzq5Dx7jz7ZWsSznMPcPbMX54O6pVs0CXJSIiImeZ30KYc26PmSWbWQfn3BZgOLDR+28M8JT364f+qqG8WbJ9P3fNWEV2XgGTb+zNBV2aBLokERERCRB/zxN2NzDDzEKB7cBYPF2gs83sFmAHcLWfawg45xxTf0jkiU830TKiFpNvjKNtozqBLktEREQCyK8hzDm3Bogr4q7h/nze8uRYTj4T565n7uqdnN+5MX+/ugdhNUICXZaIiIgEmGbM96PkA0e58+2VbNx9hPvOb8/vzmur8V8iIiICKIT5zQ/xafxu5iryChyvj4ljWMciZ+IQERGRKkohrIw555iy8Cee/GwTbSLrMPmmOFo1rB3oskRERKScUQgrQ0dz8nhgzno+XruLkV2b8LerelCnut5iERER+SUlhDKStP8ot7+1gi1705lwYQf+Z2gbzDT+S0RERIqmEFYGvtuayt3vrPZMRXFzH4Z2aBTokkRERKScUwg7A845/vntdv72+WbaNw7jXzf2pmWExn+JiIhI6RTCTlNmdh4T3l/Lp+v3cEn3pjxzZXdqhertFBEREd8oNZyGxLRMbn9rBfH7Mph4UUduG9xa479ERETklCiEnaL5m/dxz7urCapmTB/Xj3PaNQx0SSIiIlIBKYT5qKDA8Y/58fz9q610alKXf93Ym6gGtQJdloiIiFRQCmE+SM/K5Q+z1/LFxr38umcznryiOzVDgwJdloiIiFRgCmGlSEjN4PbpK0jcf5RHL+nMuEExGv8lIiIiZ0whrARfbtzLfbPWEBJcjbdu6cvANhr/JSIiImVDIawIBQWOSV9vY9LX2+jWvB7/vLE3zcNrBrosERERqUQUwk5yJCuX37+7hq837+OKXs356+XdqBGi8V8iIiJSthTCCtm2N5073lpJ0oGj/PnSLtw0oKXGf4mIiIhfKIR5zduwmz/MXkvN0CBm3tafvv/f3v2G6lnXcRx/fzwTmgpTM8ZyqyM0FMvmZA8soQdaEBkZ9EAjRcInSeiKKO1JjyIkImwlhVm2cOiDZRo9GMqUChz9szX3JxBs2eqsnRGzFmFm3x7cv9Fhbe6c222/Xfd5v+Dm/l2/c3Odz8UPrvO9f9fvXNclF/aOJEmSJphFGPD9bXv5wuO7WLPqfL5181WsWOb6L0mSdGqlqnpnOKEks8AfeueYABcBB3uH0Ngcv+FzDIfPMRy+0zGGb62qN53oQ4MownRyJPlVVa3rnUPjcfyGzzEcPsdw+M6kMTyrdwBJkqTFyCJMkiSpA4uwxeX+3gH0ujh+w+cYDp9jOHxnzBi6JkySJKkDZ8IkSZI6sOZUnt8AAARUSURBVAiTJEnqwCJswiVZleTpJLuT7EqyvncmjSfJVJLfJPlx7yxauCTnJ9mc5HdJ9iR5V+9Mmr8kn27n0J1JHk7yht6Z9NqSfDfJgSQ75/RdmOTJJM+39wt6ZrQIm3z/Bj5TVZcDVwOfTHJ550waz3pgT+8QGtvXgC1VdRmwBsdyMJJcDNwJrKuqdwBTwE19U2kevge8/6i+u4GtVbUa2Nq2u7EIm3BVNVNVz7b23xmd+C/um0oLlWQlcD3wQO8sWrgky4D3AN8BqKp/VdWhvqm0QEuApUmWAOcAf+6cRydQVT8F/npU9w3AxtbeCHz4tIY6ikXYIpJkGlgL/LxvEo3hXuBzwH96B9FYLgFmgQfbJeUHkpzbO5Tmp6r+BHwFeBGYAV6qqif6ptKYllfVTGvvB5b3DGMRtkgkOQ/4AfCpqvpb7zyavyQfBA5U1a97Z9HYlgBXAd+sqrXAP+h8GUTz19YN3cComH4zcG6Sm/um0utVo3t0db1Pl0XYIpDkbEYF2KaqerR3Hi3YNcCHkuwFHgGuTfJQ30haoH3Avqo6Mgu9mVFRpmF4L/D7qpqtqleAR4F3d86k8fwlyQqA9n6gZxiLsAmXJIzWoeypqq/2zqOFq6rPV9XKqppmtBj4qaryW/iAVNV+4I9JLm1d1wG7O0bSwrwIXJ3knHZOvQ7/sWKofgTc2tq3Ao93zGIRtghcA9zCaPZke3t9oHcoaRG6A9iUZAdwJfClznk0T20GczPwLPAco7+dZ8yjb3RsSR4GtgGXJtmX5DbgHuB9SZ5nNMN5T9eMPrZIkiTp9HMmTJIkqQOLMEmSpA4swiRJkjqwCJMkSerAIkySJKkDizBJg5Tk1Tm3Xdme5KTdgT7JdJKdJ2t/knQsS3oHkKQx/bOqruwdQpLG5UyYpImSZG+SLyd5Lskvkryt9U8neSrJjiRbk7yl9S9P8sMkv22vI4+jmUry7SS7kjyRZGn7/J1Jdrf9PNLpMCVNAIswSUO19KjLkTfO+dlLVXUF8A3g3tb3dWBjVb0T2ARsaP0bgJ9U1RpGz3Pc1fpXA/dV1duBQ8BHWv/dwNq2n0+cqoOTNPm8Y76kQUpyuKrOO0b/XuDaqnqhPbx+f1W9MclBYEVVvdL6Z6rqoiSzwMqqennOPqaBJ6tqddu+Czi7qr6YZAtwGHgMeKyqDp/iQ5U0oZwJkzSJ6jjthXh5TvtV/reG9nrgPkazZr9M4tpaSWOxCJM0iW6c876ttZ8BbmrtjwE/a+2twO0ASaaSLDveTpOcBayqqqeBu4BlwP/NxknSfPgNTtJQLU2yfc72lqo6cpuKC5LsYDSb9dHWdwfwYJLPArPAx1v/euD+JLcxmvG6HZg5zu+cAh5qhVqADVV16KQdkaRFxTVhkiZKWxO2rqoO9s4iSa/Fy5GSJEkdOBMmSZLUgTNhkiRJHViESZIkdWARJkmS1IFFmCRJUgcWYZIkSR38F4B/hdx4f/bwAAAAAElFTkSuQmCC\n",
      "text/plain": [
       "<Figure size 720x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "show_results(\n",
    "    {k:v for k, v in resistance_reports.items() if 'CNN' in k}, \n",
    "    title=\"Resistance of CNNs with CNN protection\"\n",
    ")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 6.3 Non-NN models\n",
    "\n",
    "We now test many models from the sklearn library"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Data preparation"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [],
   "source": [
    "transform = BaseNet(quadratic_model)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [],
   "source": [
    "def get_input_onehot_labels(dataset, label=\"font\", one_hot=True):\n",
    "    data_input = dataset.tensors[0]\n",
    "    label_idx = {'char': 0, 'font': 1}[label]\n",
    "    label_size = {'char': N_CHARS, 'font': N_FONTS}[label]\n",
    "    labels = dataset.tensors[1][:, label_idx].view(-1, 1)\n",
    "    \n",
    "    data_label_onehot = torch.zeros(len(labels), label_size)\n",
    "    data_label_onehot.scatter_(1, labels, 1)\n",
    "    \n",
    "    return data_input, labels, data_label_onehot\n",
    "    "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Get dataset and transform in one hot vector"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Training set 60000 items\n",
      "Testing set  10000 items\n"
     ]
    }
   ],
   "source": [
    "train_dataset, test_dataset = collateral.data.get_datasets(None, None)\n",
    "train_input, train_label, train_label_one_hot = get_input_onehot_labels(train_dataset, label=\"font\")\n",
    "test_input, test_label, test_label_one_hot = get_input_onehot_labels(test_dataset, label=\"font\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Apply the quadratic model transformation"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [],
   "source": [
    "train_input = transform.quad(train_input).detach().numpy()\n",
    "test_input = transform.quad(test_input).detach().numpy()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [],
   "source": [
    "train_label = train_label.reshape(-1)\n",
    "test_label = test_label.reshape(-1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "((60000, 10), torch.Size([60000]))"
      ]
     },
     "execution_count": 22,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "train_input.shape, train_label.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {},
   "outputs": [],
   "source": [
    "ALL = train_input.shape[0]\n",
    "CPOWER = 'LOW'"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn import linear_model\n",
    "from sklearn import kernel_ridge\n",
    "from sklearn import svm"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {},
   "outputs": [],
   "source": [
    "def evaluate_sklearn(reg, one_hot=True, limit=int(10e10)):\n",
    "    train_labels = {True: train_label_one_hot, False: train_label}[one_hot]\n",
    "    reg.fit(train_input[:limit], train_labels[:limit].detach().numpy()) \n",
    "    output = reg.predict(test_input)\n",
    "    if one_hot:\n",
    "        pred = torch.tensor(output).argmax(1, keepdim=True)\n",
    "    else:\n",
    "        if isinstance(output, list):\n",
    "            pred = torch.tensor(list(map(round, output))).long().view(-1, 1)\n",
    "        else:\n",
    "            pred = torch.tensor(np.round(output)).long().view(-1, 1)\n",
    "    y = test_label.view_as(pred)\n",
    "    acc = pred.eq(y).sum().item() / len(pred)\n",
    "    return acc"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Linear models"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.502\n"
     ]
    }
   ],
   "source": [
    "reg = linear_model.Ridge(alpha=.9)\n",
    "acc = evaluate_sklearn(reg)\n",
    "print(acc)\n",
    "resistance_reports['linear model Ridge'] = acc * 100"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.502\n"
     ]
    }
   ],
   "source": [
    "reg = linear_model.RidgeCV(alphas=[0.01, 0.1, 1.0, 10.0, 100], cv=5)\n",
    "acc = evaluate_sklearn(reg)\n",
    "print(acc)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.5005\n"
     ]
    }
   ],
   "source": [
    "reg = linear_model.Lasso(alpha=0.1)\n",
    "acc = evaluate_sklearn(reg)\n",
    "print(acc)\n",
    "resistance_reports['linear model Lasso'] = acc * 100"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.5005\n"
     ]
    }
   ],
   "source": [
    "reg = linear_model.LassoLars(alpha=0.1)\n",
    "acc = evaluate_sklearn(reg)\n",
    "print(acc)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.5032\n"
     ]
    }
   ],
   "source": [
    "reg = linear_model.BayesianRidge()\n",
    "acc = evaluate_sklearn(reg, one_hot=False)\n",
    "print(acc)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.5019\n"
     ]
    }
   ],
   "source": [
    "reg = linear_model.LogisticRegression(random_state=0, solver='lbfgs',\n",
    "                         multi_class='multinomial', max_iter=1000)\n",
    "acc = evaluate_sklearn(reg, one_hot=False)\n",
    "print(acc)\n",
    "resistance_reports['logistic regression'] = acc * 100"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.5031\n"
     ]
    }
   ],
   "source": [
    "reg = linear_model.LogisticRegressionCV(cv=5, random_state=0, multi_class='multinomial')\n",
    "acc = evaluate_sklearn(reg, one_hot=False)\n",
    "print(acc)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Quadratic Discriminant Analysis"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn.discriminant_analysis import QuadraticDiscriminantAnalysis"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.5465\n"
     ]
    }
   ],
   "source": [
    "clf = QuadraticDiscriminantAnalysis()\n",
    "acc = evaluate_sklearn(clf, one_hot=False)\n",
    "print(acc)\n",
    "resistance_reports['Quadratic Discriminant Analysis'] = acc * 100"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Kernel Ridge"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/Users/tryffel/code/env/pysyft/lib/python3.7/site-packages/sklearn/linear_model/ridge.py:165: LinAlgWarning: Ill-conditioned matrix (rcond=8.07553e-11): result may not be accurate.\n",
      "  overwrite_a=False)\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.5053\n"
     ]
    }
   ],
   "source": [
    "clf = kernel_ridge.KernelRidge(alpha=1.0)\n",
    "limit = {'LOW': 15000, 'MEDIUM': ALL, 'HIGH': ALL}\n",
    "acc = evaluate_sklearn(clf, one_hot=False, limit=limit[CPOWER])\n",
    "print(acc)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### SVM"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 36,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.5108\n"
     ]
    }
   ],
   "source": [
    "clf = svm.SVC(gamma='scale', kernel='linear')\n",
    "limit = {'LOW': 5000, 'MEDIUM': 15000, 'HIGH': ALL}\n",
    "acc = evaluate_sklearn(clf, one_hot=False, limit=limit[CPOWER])\n",
    "\n",
    "print(acc)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.5574\n"
     ]
    }
   ],
   "source": [
    "clf = svm.SVC(gamma='scale', kernel='rbf')\n",
    "limit = {'LOW': 20000, 'MEDIUM': ALL, 'HIGH': ALL}\n",
    "acc = evaluate_sklearn(clf, one_hot=False, limit=limit[CPOWER])\n",
    "print(acc)\n",
    "resistance_reports['SVM (rbf)'] = acc * 100"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 38,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.5065\n"
     ]
    }
   ],
   "source": [
    "clf = svm.SVC(gamma='scale', kernel='poly', coef0=0)\n",
    "limit = {'LOW': 20000, 'MEDIUM': ALL, 'HIGH': ALL}\n",
    "acc = evaluate_sklearn(clf, one_hot=False, limit=limit[CPOWER])\n",
    "print(acc)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 39,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.4882\n"
     ]
    }
   ],
   "source": [
    "clf = svm.SVC(gamma='scale', kernel='sigmoid', coef0=0)\n",
    "limit = {'LOW': 20000, 'MEDIUM': ALL, 'HIGH': ALL}\n",
    "acc = evaluate_sklearn(clf, one_hot=False, limit=limit[CPOWER])\n",
    "print(acc)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Stochastic gradient Descent"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 40,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.5003\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/Users/tryffel/code/env/pysyft/lib/python3.7/site-packages/sklearn/linear_model/stochastic_gradient.py:166: FutureWarning: max_iter and tol parameters have been added in SGDClassifier in 0.19. If both are left unset, they default to max_iter=5 and tol=None. If tol is not None, max_iter defaults to max_iter=1000. From 0.21, default max_iter will be 1000, and default tol will be 1e-3.\n",
      "  FutureWarning)\n"
     ]
    }
   ],
   "source": [
    "clf = linear_model.SGDClassifier(loss=\"hinge\", penalty=\"l2\")\n",
    "acc = evaluate_sklearn(clf, one_hot=False)\n",
    "print(acc)\n",
    "resistance_reports['SGDClassifier'] = acc * 100"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 41,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.4919\n"
     ]
    }
   ],
   "source": [
    "clf = linear_model.SGDClassifier(max_iter=1000, tol=1e-3, loss=\"log\")\n",
    "acc = evaluate_sklearn(clf, one_hot=False)\n",
    "print(acc)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### K-Neighbors Classifier"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 42,
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn.neighbors import KNeighborsClassifier"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 43,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.8018\n"
     ]
    }
   ],
   "source": [
    "clf = KNeighborsClassifier(n_neighbors=7)\n",
    "acc = evaluate_sklearn(clf, one_hot=False)\n",
    "print(acc)\n",
    "resistance_reports['K-Neighbors Classifier'] = acc * 100"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Gaussian process (not scalable)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 44,
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn.gaussian_process import GaussianProcessClassifier\n",
    "from sklearn.gaussian_process.kernels import RBF"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 45,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.6193\n"
     ]
    }
   ],
   "source": [
    "kernel = 1.0 * RBF(1.0)\n",
    "gpc = GaussianProcessClassifier(kernel=kernel,\n",
    "        random_state=0)\n",
    "limit = {'LOW': 1700, 'MEDIUM': 5000, 'HIGH': 20000}\n",
    "acc = evaluate_sklearn(gpc, one_hot=False, limit=limit[CPOWER])\n",
    "print(acc)\n",
    "resistance_reports['Gaussian process'] = acc * 100"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Naive Bayes"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 46,
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn.naive_bayes import GaussianNB"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 47,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.5326\n"
     ]
    }
   ],
   "source": [
    "\n",
    "gnb = GaussianNB()\n",
    "acc = evaluate_sklearn(gnb, one_hot=False)\n",
    "print(acc)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Decision Tree Classifier"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 48,
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn.tree import DecisionTreeClassifier"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 49,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.5591\n"
     ]
    }
   ],
   "source": [
    "\n",
    "clf = DecisionTreeClassifier(max_depth=5)\n",
    "acc = evaluate_sklearn(clf, one_hot=False)\n",
    "print(acc)\n",
    "\n",
    "resistance_reports['Decision Tree Classifier'] = acc * 100"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Ensemble methods"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 50,
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn.ensemble import RandomForestClassifier\n",
    "from sklearn.ensemble import AdaBoostClassifier\n",
    "from sklearn.ensemble import GradientBoostingClassifier"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 51,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.8046\n"
     ]
    }
   ],
   "source": [
    "clf = RandomForestClassifier(max_depth=30, n_estimators=100, max_features=4)\n",
    "acc = evaluate_sklearn(clf, one_hot=False)\n",
    "print(acc)\n",
    "resistance_reports['Random Forest Classifier'] = acc * 100"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 52,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.5659\n"
     ]
    }
   ],
   "source": [
    "clf = AdaBoostClassifier(n_estimators=200, learning_rate=1.)\n",
    "acc = evaluate_sklearn(clf, one_hot=False)\n",
    "print(acc)\n",
    "resistance_reports['AdaBoost Classifier'] = acc * 100"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 53,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.6891\n"
     ]
    }
   ],
   "source": [
    "clf = GradientBoostingClassifier(n_estimators=10, learning_rate=1.0,\n",
    "    max_depth=10, random_state=0)\n",
    "acc = evaluate_sklearn(clf, one_hot=False)\n",
    "print(acc)\n",
    "resistance_reports['Gradient Boosting Classifier'] = acc * 100"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### (sklearn) Neural Network"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 54,
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn.neural_network import MLPClassifier"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 55,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.5487\n"
     ]
    }
   ],
   "source": [
    "clf = MLPClassifier(solver='lbfgs', alpha=1,\n",
    "                    hidden_layer_sizes=(10, 40, 10), random_state=1)\n",
    "acc = evaluate_sklearn(clf, one_hot=False)\n",
    "print(acc)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Summary"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 56,
   "metadata": {},
   "outputs": [],
   "source": [
    "def print_table(results):\n",
    "    n_cols = 2\n",
    "    title_length = 0\n",
    "    for title in results.keys():\n",
    "        title_length = max(title_length, len(title))\n",
    "    result_length = 8\n",
    "    table_length = 1 + n_cols + title_length + result_length\n",
    "    sep = '+'.join(\n",
    "        ['', '-'*title_length, '-'*result_length, '']\n",
    "    )\n",
    "    for title, serie in results.items():\n",
    "        print(sep)\n",
    "        title = (title + ' '*title_length)[:title_length]\n",
    "        if isinstance(serie, list):\n",
    "            result = round(np.mean(serie[-4:]), 2)\n",
    "        else:\n",
    "            result = round(serie, 2)\n",
    "        result = (' ' +str(result) + '%' + ' '*result_length)[:result_length]\n",
    "        line = '|'.join(\n",
    "            ['', title, result, '']\n",
    "        )\n",
    "        print(line)\n",
    "    print(sep)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 57,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "+---------------------------------------+--------+\n",
      "|Collateral task with net 8:64:32:16:8:2| 59.42% |\n",
      "+---------------------------------------+--------+\n",
      "|Collateral task with net 8:32:16:8:2   | 63.96% |\n",
      "+---------------------------------------+--------+\n",
      "|Collateral task with net 8:24:12:2     | 65.53% |\n",
      "+---------------------------------------+--------+\n",
      "|Collateral task with net 8:64:2        | 68.38% |\n",
      "+---------------------------------------+--------+\n",
      "|Collateral task with net 8:32:2        | 66.31% |\n",
      "+---------------------------------------+--------+\n",
      "|Collateral task with net 8:16:2        | 64.11% |\n",
      "+---------------------------------------+--------+\n",
      "|Collateral task with CNN 0             | 69.78% |\n",
      "+---------------------------------------+--------+\n",
      "|linear model Ridge                     | 50.2%  |\n",
      "+---------------------------------------+--------+\n",
      "|linear model Lasso                     | 50.05% |\n",
      "+---------------------------------------+--------+\n",
      "|logistic regression                    | 50.19% |\n",
      "+---------------------------------------+--------+\n",
      "|Quadratic Discriminant Analysis        | 54.65% |\n",
      "+---------------------------------------+--------+\n",
      "|SVM (rbf)                              | 55.74% |\n",
      "+---------------------------------------+--------+\n",
      "|SGDClassifier                          | 50.03% |\n",
      "+---------------------------------------+--------+\n",
      "|K-Neighbors Classifier                 | 80.18% |\n",
      "+---------------------------------------+--------+\n",
      "|Gaussian process                       | 61.93% |\n",
      "+---------------------------------------+--------+\n",
      "|Decision Tree Classifier               | 55.91% |\n",
      "+---------------------------------------+--------+\n",
      "|Random Forest Classifier               | 80.46% |\n",
      "+---------------------------------------+--------+\n",
      "|AdaBoost Classifier                    | 56.59% |\n",
      "+---------------------------------------+--------+\n",
      "|Gradient Boosting Classifier           | 68.91% |\n",
      "+---------------------------------------+--------+\n"
     ]
    }
   ],
   "source": [
    "print_table(resistance_reports)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Compative results\n",
    "\n",
    "If you kept sabotage epochs at 10 you might have high collateral accuracies.\n",
    "Here is the table you would get after a long resistance phase:"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "```\n",
    "+---------------------------------------+--------+\n",
    "|Collateral task with net 8:64:32:16:8:2| 55.52% |\n",
    "+---------------------------------------+--------+\n",
    "|Collateral task with net 8:32:16:8:2   | 54.71% |\n",
    "+---------------------------------------+--------+\n",
    "|Collateral task with net 8:24:12:2     | 54.54% |\n",
    "+---------------------------------------+--------+\n",
    "|Collateral task with net 8:64:2        | 55.16% |\n",
    "+---------------------------------------+--------+\n",
    "|Collateral task with net 8:32:2        | 54.53% |\n",
    "+---------------------------------------+--------+\n",
    "|Collateral task with net 8:16:2        | 54.57% |\n",
    "+---------------------------------------+--------+\n",
    "|linear model Ridge                     | 49.66% |\n",
    "+---------------------------------------+--------+\n",
    "|linear model Lasso                     | 50.05% |\n",
    "+---------------------------------------+--------+\n",
    "|logistic regression                    | 49.69% |\n",
    "+---------------------------------------+--------+\n",
    "|Quadratic Discriminant Analysis        | 50.25% |\n",
    "+---------------------------------------+--------+\n",
    "|SVM (rbf)                              | 53.2%  |\n",
    "+---------------------------------------+--------+\n",
    "|SGDClassifier                          | 49.96% |\n",
    "+---------------------------------------+--------+\n",
    "|K-Neighbors Classifier                 | 58.22% | *\n",
    "+---------------------------------------+--------+\n",
    "|Gaussian process                       | 50.05% |\n",
    "+---------------------------------------+--------+\n",
    "|Decision Tree Classifier               | 52.43% |\n",
    "+---------------------------------------+--------+\n",
    "|Random Forest Classifier               | 59.2%  | *\n",
    "+---------------------------------------+--------+\n",
    "|AdaBoost Classifier                    | 54.01% |\n",
    "+---------------------------------------+--------+\n",
    "|Gradient Boosting Classifier           | 57.04% | *\n",
    "+---------------------------------------+--------+\n",
    "```"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Conclusion\n",
    "\n",
    "The long term resistance has proved to be very effective and the collateral accuracy we get is even lower than previous results of Part 6!\n",
    "For Neural Networks the results are very good as we don't exceed 55-56%. For other models likes K-Neighbors or Random Forest, we observe also a significant drop as we get under 60%\n",
    "\n",
    "> This is only true if you set at least 50 epochs to build resistance"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "You can find extended results in Part 8 bis, but they are a bit long to compute!"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.7.0"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}